Most bad agency hires happen because the buyer takes sales calls first, works out what they need second and reads the contract last. This guide reverses that order: it is the procedure we would follow if we were hiring, and you can run it without anyone's help, including ours.
To choose an AI automation agency, write a one-page project brief, ask every candidate the same nine questions (scope, ownership, testing, data, human oversight, support, pricing, references, handover), score the answers on a weighted scorecard and prove the winner with a paid two to four week pilot. Pick the agency whose answers are specific, whose build you will own and who can show how it fails safely.
Disclosure: Six Figure Consulting is itself an AI automation agency, so we have an interest in this topic. We wrote this guide so that you can evaluate any agency, including us, with the same scorecard. If a different agency scores higher for your project, hire them. Nothing here is legal advice; have a qualified lawyer review any contract before you sign.
If the term is new to you, read what an AI agent is first. This guide is for B2B founders, consultants and agency owners in the UAE, the US and the wider GCC about to pay someone to build AI automation and who want to avoid paying for a demo that never survives real data.
What you will have, and what you need first
At the end: a one-page brief, written answers to nine questions scored on one weighted sheet, two reference calls per finalist, a paid pilot with success and exit criteria, a contract checklist for your lawyer and a signed pilot, ideally within three weeks. Before you start: one process with an owner on your side, two weeks of measurements of it (volume and time per item), two people to score independently, a lawyer who knows your jurisdiction and sample data you may share, anonymised if needed.
What you are actually buying
An AI automation agency sells a working system that does a repeatable task for you: reading and routing enquiries, researching prospects, drafting replies, updating records, producing reports. The valuable part is rarely the model. It is the design around it: what triggers the system, what it may do, how its output is checked, what happens when it is wrong and who maintains it.
Anthropic's engineering team separates workflows (paths written in advance) from agents (the model decides its own steps) and advises the simplest solution that works, because autonomy adds cost and lets errors compound. It also warns that frameworks can hide the prompts and responses underneath, which makes debugging harder. (Source: Anthropic, "Building effective agents", accessed 6 November 2026.) For a buyer that becomes a practical test: a good agency will tell you when a plain workflow is the better answer, and will show you the prompts and logic inside what they build.
If you are unsure whether you need a builder or a strategist first, our comparison of an AI consulting company vs an AI automation agency covers that decision. If your project is outbound sales, AI SDR vs human SDR shows the cost and handoff questions to settle before you brief anyone.
The six steps
Step 1: Write the project brief
Input: your two weeks of measurements and the process owner. Do: fill every field of the template below. Where you do not know an answer, write "unknown" rather than guessing, because an agency's questions about the gaps tell you a lot about them. Output: one page you can send to every candidate. Check: hand it to a colleague who was not in your planning. If they can explain the problem, the finish line and the biggest risk back to you in two minutes, it is ready.
Asset 1: Project brief templatePROJECT BRIEF: <PROJECT_NAME>
Owner: <NAME, ROLE> Date: <DATE>
1. PROBLEM (one or two sentences)
<PROBLEM_STATEMENT>
2. CURRENT PROCESS (each step: owner, minutes per item)
<STEP> | <ROLE> | <MINUTES>
Volume per month: <N> Busiest week: <N>
Edge cases: <LIST>
3. SYSTEMS
Inputs from: <FORMS, EMAIL, CRM, SHEET>
Outputs into: <CRM, HELPDESK, REPORT>
Accounts we pay for: <TOOLS> Admin held by: <NAME>
4. SUCCESS METRIC
Metric: <METRIC>
Baseline (measured, and how): <VALUE>
Good enough to keep: <TARGET> Stop if: <THRESHOLD>
5. CONSTRAINTS
Must never happen: <E.G. UNAPPROVED CUSTOMER MESSAGES>
Tone and languages: <LIST>
Regulatory or contractual limits: <LIST OR "TO CHECK">
6. DATA
Personal data: <YES/NO, WHICH>
Confidential client or financial data: <YES/NO>
Must stay in a country or system: <LIST>
Pilot sample data: <ANONYMISED/REAL, WHO APPROVES>
7. PEOPLE
Outcome owner: <NAME> Technical contact: <NAME>
Signs the contract: <NAME>
8. BUDGET
Pilot: <RANGE> Ongoing monthly incl. running costs: <RANGE>
Preferred model: <PROJECT/RETAINER/USAGE/OPEN>
9. TIMELINE
Pilot start: <DATE> Decision on full build: <DATE>
Hard deadlines: <DATE AND REASON>
Step 2: Shortlist three agencies and send the same email
Input: the brief. Do: pick three candidates (look for named people, case studies and a description of how they work) and send each the email below with the brief attached. Before reading any reply, write down your scoring weights and who will score. Output: three comparable proposals. Check: the questions that come back show how each agency thinks; one that asks nothing is pitching a standard package.
Asset 2: Request for proposals emailSubject: Proposal request: <PROJECT_NAME> (response by <DATE>)
Hi <NAME>,
We want to automate <ONE_LINE_PROCESS>. The attached brief has
the problem, volumes, systems, constraints and timeline. We are
comparing three agencies on the same criteria. Please reply with:
1. Your understanding of the scope, and what you would exclude.
2. How you would measure success against our baseline.
3. Who would work on this, and their role.
4. Where it would run and who would own each account.
5. How you would test it before it touches a customer.
6. Every tool and subcontractor that would see our data, with
region and retention terms.
7. How a person approves, pauses or overrides it, and what
happens when it is wrong.
8. A pilot proposal: duration, fixed fee, deliverables, exit
criteria.
9. Support options, and what is billed separately, including
running costs.
10. Two references we can call, one of similar size.
11. A sample handover pack (anonymised) or its contents.
We will hold 45 minute calls in the week of <DATE>, decide by
<DATE> and start a paid pilot by <DATE>. Tell us if anything in
the brief looks wrong.
<YOUR_NAME>, <ROLE>, <COMPANY>
Step 3: Interview, score and verify
Input: the written proposals. Do: hold one 45 minute call per agency, ask the nine questions below in the same order and take notes. Score on the scorecard, apply the hard rule, check red flags, then call two references per finalist. Output: a scored sheet and one or two finalists. Check: every score above 3 rests on something in writing or something you saw demonstrated.
The 9 questions
Ask all nine, in the same order. The example answers are illustrative wording, not quotes from any real agency. Use the probe when the first answer is polished but vague.
| Question | Strong answer sounds like | Weak answer sounds like | Probe |
|---|---|---|---|
| 1. Scope: which exact process will you automate, and what does done look like? | Names the process and inputs, a finish line at an accuracy you set, and what is out of scope | "We will transform your customer experience with AI agents." | "Name two things you would not automate here, and why." |
| 2. Ownership: who owns the accounts, prompts, workflows and data? | Everything is built in accounts you create and administer; revocable invited access; prompts, exports and docs are yours from day one | "It runs on our platform." Or "the workflows are our IP, but you can license them." | "If we ended tomorrow, list what we keep, in what format, within how many days. Does any tool need a licence in your name?" |
| 3. Testing: how will we know it works before it touches a customer? | An agreed test set of your real examples including awkward ones, scored against your criteria, re-run after every change, then shadow mode (drafting, not sending) with results shared | "We test it internally and it works very well." | "Show me a test report from a past project with the failures left in." |
| 4. Data: what will the system see, where does it go and who can access it? | A tool-by-tool map with each provider's region and retention and training terms, only the fields the task needs, credentials in your vault, logged access, a DPA for personal data | "Your data is completely safe, we use enterprise AI." | "Which subcontractors see our data, where is each processed, is any used for training? In writing, please." |
| 5. Oversight: what happens when the AI gets it wrong? | Low-risk actions run automatically; customer-facing, financial or irreversible ones need approval; low-confidence cases go to a named person; logs and a pause switch you control | "It rarely makes mistakes." Or "no need for you to be involved." | "Describe the worst failure on a past project and what you changed." |
| 6. Support: what happens after launch? | A hypercare period (say 30 days), then a plan with response times per severity, or documentation to run it yourself; clear bug versus change rules | "We are always here for you." No times, scope or price. | "Give two examples each of a bug (support) and a change request (billed)." |
| 7. Pricing: what will I pay, for what, and what is excluded? | Fixed pilot fee; running costs billed by providers to your accounts, estimated from the pilot; support cancellable with notice; written quotes for changes | "A custom package once you are on board," or one large figure with lock-in and auto-renewal | "What is the monthly running cost once live, and who sees the invoices?" |
| 8. References and team: who does the work, and who can vouch for it? | Named builders, two callable references (one similar), honest limits on confidential work | "Our team of experts will handle it." Anonymous testimonials only. | "What would a past client say you could have done better?" |
| 9. Handover: what exactly do we receive at the end? | A pack with a workflow diagram, account and integration list, annotated prompts, test set and results, operating guide, known limits and credential register, plus a sign-off test | "We will walk you through it at the end." | "May I see an anonymised handover pack?" |
Why these answers carry weight. If the system runs in the agency's accounts you are renting it and may lose workflows, prompts and history when you part. On testing, Anthropic's evaluation documentation advises specific, measurable success criteria, tests that mirror the real task including edge cases, and automated grading where possible, noting that hand grading is higher quality but does not scale. (Source: Claude developer documentation, "Define success criteria and build evaluations", accessed 6 November 2026.) In our setup we build the whole flow on a sample of 10 to 20 items first, read every output, estimate the cost of the next batch and only then scale. On data, providers publish location and retention controls; Anthropic documents that customers can set where model inference runs and where data is stored at rest, with options that differ by platform and model, and zero data retention only for eligible features. (Source: Claude developer documentation, "Data residency", accessed 6 November 2026.) On oversight, Anthropic's agent guidance stresses human checkpoints, stopping conditions and guardrails, since errors compound when a system acts autonomously. (Source: Anthropic, "Building effective agents", accessed 6 November 2026.)
The scorecard
The last pitch you heard usually wins unless you score on paper. Each criterion maps to one question. The weights are our judgment, tuned to the hardest-to-reverse risks (ownership, data, oversight). Change them if your situation differs, but before you see any proposal.
Asset 3: Weighted scorecard| Criterion (question) | Weight | 1 (poor) | 3 (acceptable) | 5 (excellent) |
|---|---|---|---|---|
| 1. Scope clarity | 12 | Vague promises, no exclusions | Process named, finish line partly defined | Written scope, finish line, exclusions, metric |
| 2. Ownership of accounts and IP | 15 | Runs in agency accounts; agency owns prompts and flows | Mostly your accounts; some lock-in | Your accounts; you own prompts, flows and data from day one |
| 3. Testing | 12 | "Tested internally" | Manual sample testing, some criteria | Agreed test set with edge cases, shared results, shadow mode |
| 4. Data handling | 14 | Reassurance only | Tool list, some gaps | Tool-by-tool map, regions and retention, DPA, logged access |
| 5. Human oversight | 11 | "It rarely fails"; no approvals or logs | Some approvals, basic logs | Risk-tiered approvals, logs, pause switch, named escalation |
| 6. Support | 9 | None defined | Offered, response times loose | Hypercare, severity levels, bug versus change rules |
| 7. Pricing | 9 | One opaque figure, long lock-in | Itemised fee, running costs partly clear | Itemised, running costs separate, cancel with notice |
| 8. References and team | 8 | None; unnamed team | Case studies; one call possible | Named team, two callable references |
| 9. Handover | 10 | Verbal walkthrough only | Some docs, no acceptance test | Named pack, sample shown, sign-off test |
Score each criterion from 1 to 5 (use 2 and 4 for in-between answers), multiply by the weight, add the nine results and divide by 5. The weights sum to 100 and the top score is 5, so the maximum is 100.
Total = (S1 x 12 + S2 x 15 + S3 x 12 + S4 x 14 + S5 x 11
+ S6 x 9 + S7 x 9 + S8 x 8 + S9 x 10) / 5
| Total | Reading | Action |
|---|---|---|
| 80 to 100 | Strong on the risks that matter | Proceed to references and a paid pilot |
| 65 to 79 | Workable with named gaps | Ask for the gaps to be fixed in writing, then pilot |
| 50 to 64 | Weak on several fronts | Pass unless the project is tiny and the gaps cost little |
| Below 50 | Not ready for your project | Pass |
These bands are our rule of thumb, not an industry standard. Add one hard rule: a score of 1 on ownership, data handling or human oversight disqualifies the agency regardless of total, because nothing else compensates for a build you cannot take away, data you cannot trace or an unsupervised system.
- Have two people score independently, then average; discuss any gap of two points or more.
- Score only from evidence in writing or seen demonstrated. A promise on a call counts as 3 at most.
- Read the pattern as well as the total: a 4 average with one 1 is riskier than a steady 3.
- A gap under 5 points is a tie; break it with references and running cost. Nudge your two least certain scores by one point: if the ranking changes, test that point in the pilot.
Red flags and what to ask next
Asset 4: Red flag table| Signal | What it usually means | What to ask next |
|---|---|---|
| Guaranteed leads, savings or revenue before seeing your process, or only perfect demo examples | Selling a promise, not a plan | "What would have to be true for this outcome, and what do you do if it is not?" |
| Runs only inside their accounts or platform | Lock-in; you rent rather than own | "Can this run in accounts I own? What do I keep on exit?" |
| Refuses to show prompts or logic | Black box nobody else can maintain | "Will prompts and workflow exports be in the handover pack?" |
| Vague answers on data and subcontractors | No data map; unknown exposure | "List every tool that sees our data, with region and retention terms." |
Reference calls
Ask for two references per finalist, one resembling your project. Keep each call to 20 minutes, say you will keep it confidential and that you want to hear what went wrong as well as what went well.
Asset 5: Reference-call script (8 questions)- What did you hire them to do, and when?
- Was the scope clear, and did it change? Who initiated the changes?
- Who actually did the work, and did that match who you met in sales?
- How did they test before go-live? Did you see results?
- What went wrong after launch, how fast did they respond and what changed?
- Does it run in your own accounts? Could you move it without them?
- Were the costs as expected, including running costs, and was the handover good?
- Would you hire them again, and what should I ask them?
Write the answers against the scorecard rows, and be cautious if every reference is glowing but unspecific.
Step 4: Design a paid pilot
Input: your finalist and the baseline from your brief. Do: agree a small, paid, time-boxed build of one slice of the real project, with written success and exit criteria, run in your accounts. Output: a one-page pilot agreement. Check: a stranger could read it and tell, on the last day, whether the pilot passed. An agency that welcomes a pilot is confident in its process.
First, understand how each pricing model shapes the agency's behaviour. We give no price ranges, because prices vary by scope, region and cost base and any number would mislead.
Asset 6: Pricing models compared| Model | Incentive it creates | Main risk for you | Fits best when |
|---|---|---|---|
| Fixed project fee | Finish efficiently and on time | Corners cut; change requests become the profit centre | Scope is clear and bounded, as in a pilot |
| Retainer | Keep you long term; stay responsive | Paying for idle capacity; vague deliverables | Ongoing improvement and support after launch |
| Usage-based | Encourage more usage | Unpredictable bills as volume grows | Volume varies and you can set caps and alerts |
| Performance-linked | Chase the measured outcome | Gaming the metric; attribution disputes; high price to cover risk | Outcome is easy to measure and mostly in the agency's control |
Two points apply to all four. Keep the agency's fee separate from running costs, which should sit in accounts you control. And if a performance-linked component is offered, define the metric, the measurement source, the period and the agency's contribution in writing first.
Pilot terms we recommend:
- Length: two to four weeks. Shorter hides problems; longer is a project in disguise.
- Scope: one process, one input channel, one output destination, a sample of real items. Everything else goes on a "later" list.
- Paid: a fixed fee, ideally half on start and half on acceptance. Unpaid trials get unpaid attention.
- Your accounts: create them at the start; it tests whether the agency really builds in your environment.
- Weekly checkpoint: 30 minutes with a shared results sheet, not slides.
| Success criterion (agree before day one) | Example measure | Evidence |
|---|---|---|
| Quality | Share of outputs your reviewer accepts against the rubric | Test set results |
| Safety | Nothing customer-facing sent without approval; no data outside agreed tools | Logs, tool list |
| Time | Median handling time versus your baseline | Log timestamps |
| Cost | Running cost per item, estimated from the pilot | Provider dashboards |
| Collaboration | Questions answered in the agreed time; checkpoints held | Message history |
| Documentation | Draft handover pack accepted by someone on your side | Pause-and-restart test |
Set the thresholds yourself, in writing, using your own baseline; do not borrow numbers from anyone else's results. Write the exit criteria too: what ends the pilot early (a data handling breach, a missed milestone beyond a grace period, quality below an agreed floor after two improvement rounds, or the agency changing the assigned people), that work done so far is yours, and that you pay only for milestones completed.
Step 5: Agree the contract and handover terms
Input: the pilot agreement and the scorecard notes. Do: turn them into written terms covering every row below and have a lawyer who knows your jurisdiction review them before you sign. This is a list of topics, not legal advice or a template; laws differ across the UAE, the US and other GCC jurisdictions. Output: a signed agreement for the pilot, with the same terms carrying into any later build. Check: the Pre-signing QA checklist later in this guide passes in full.
Asset 7: Contract and handover checklist| Topic | What to confirm in writing |
|---|---|
| Scope and deliverables | Numbered deliverables, exclusions, assumptions and acceptance tests for each |
| Ownership | You own accounts, prompts, workflows, configuration, test sets and outputs; the agency keeps only its pre-existing general tools, with a licence to you for anything embedded in the build |
| Access and credentials | Accounts in your name; agency access invited, least-privilege, revocable; credentials in your vault; integration register |
| Documentation | Pack contents, delivery date, sign-off test (someone on your side can pause, restart and read the logs) |
| Change process | How changes are requested, quoted and approved; bug versus change; response times by severity |
| Testing and acceptance | Test set, pass thresholds, shadow-mode period and who signs acceptance |
| Data processing | Data map of tools and subcontractors, regions, retention, deletion on exit, breach notification time, a DPA where personal data is involved, and whether your data may train models |
| Confidentiality | Mutual confidentiality, who may see your data, deletion at the end, whether you may be named as a client |
| Running costs | Provider fees billed directly to you or passed through at cost with invoices; caps and alerts; no unlimited promises |
| Fees, payment and staffing | Milestone schedule, renewal and cancellation terms; named lead, with notice before staff or subcontractor changes |
| Liability and warranties | What the agency warrants, limits on liability and insurance, reviewed by counsel; no guaranteed business outcomes |
| Termination and transition | Notice period, everything handed over within agreed days, transition help and its cost |
Step 6: Run the pilot and take over the build
Input: the signed agreement and accounts you created. Do: hold the weekly checkpoint, review outputs yourself against the rubric and, at the end, run the pause-and-restart test on the draft handover pack. Output: a pass, fix or stop decision against the criteria you wrote in Step 4. Check: you own all of the following.
- The working build, in accounts you own and administer.
- The prompts, workflow exports and configuration files.
- The test set, rubric and results, plus the draft handover pack and credential register.
- A cost estimate for full volume and a written list of known limitations and next steps.
Owning all of that means you can continue with this agency, switch or go in-house. That freedom is the real result of a good pilot.
When it goes wrong: troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Proposals cannot be compared, or all score 4 or 5 | Different questions asked; scoring from the pitch, not evidence | Same email and brief for all; cap unevidenced scores at 3 |
| Two finalists tie | Totals within 5 points | Break the tie with references and running cost after launch |
| Agency resists building in your accounts | Lock-in or a platform model | Score ownership 1 and apply the hard rule |
| Pilot results are vague | No baseline or thresholds written down | Pause, write the thresholds from your baseline, re-run the test set |
| Chose on the demo, or on the cheapest quote | Demos show the best case; price is easy to compare | Run their system on five of your worst examples; compare the weighted total and running cost together |
| Nobody on your side can pause the system | Handover treated as a courtesy | Make the pause-and-restart test an acceptance criterion |
A worked example (illustrative)
Fictional: every name and number is invented to show the mechanics, not real companies or typical results.
The buyer. Rana runs Cedar Ledger Advisory, a ten-person consultancy. Her brief: about 120 inbound enquiries a month, to be classified, checked against a qualification rubric, answered with a drafted reply a person approves and logged in the CRM. Baseline: about 12 minutes per enquiry, measured over two weeks. Nothing may reach a customer without approval, and the data includes names and company details.
The candidates. She sends the same email and brief to three fictional agencies: Harbor and Pine Automation (large, polished), Quillstack Studio (small, specialised) and Brightline Digital (cheapest), then scores each after one call.
| Criterion | Weight | Harbor and Pine | Quillstack | Brightline |
|---|---|---|---|---|
| 1. Scope clarity | 12 | 4 | 5 | 2 |
| 2. Ownership | 15 | 2 | 5 | 3 |
| 3. Testing | 12 | 3 | 4 | 1 |
| 4. Data handling | 14 | 3 | 4 | 2 |
| 5. Oversight | 11 | 3 | 4 | 2 |
| 6. Support | 9 | 4 | 3 | 2 |
| 7. Pricing | 9 | 3 | 4 | 5 |
| 8. References and team | 8 | 4 | 3 | 2 |
| 9. Handover | 10 | 2 | 5 | 2 |
Behind the scores: Quillstack wrote exclusions unprompted, proposed a test set from her own enquiries, listed its tools and regions and showed a sample handover pack; Harbor and Pine licenses its platform; Brightline keeps the prompts and wanted auto-send.
- Harbor and Pine: (48 + 30 + 36 + 42 + 33 + 36 + 27 + 32 + 20) / 5 = 304 / 5 = 60.8
- Quillstack: (60 + 75 + 48 + 56 + 44 + 27 + 36 + 24 + 50) / 5 = 420 / 5 = 84.0
- Brightline: (24 + 45 + 12 + 28 + 22 + 18 + 45 + 16 + 20) / 5 = 230 / 5 = 46.0
- Hard rule first. Nobody scored 1 on ownership, data or oversight, but Brightline's total is under 50. She passes despite the price.
- Read the pattern. Harbor and Pine is weak on ownership and handover (both 2). Even if it agreed to build in her accounts and both rose to 4, its total would be (304 + 30 + 20) / 5 = 70.8, still well behind Quillstack.
- Check sensitivity and cost. Quillstack's weakest points are support (3) and references (3). If both dropped to 2, its total would fall by (9 + 8) / 5 = 3.4 to 80.6, still the leader, and its running costs are separate and estimated, so the true cost is visible.
- Decide, and test the weak spots. She calls Quillstack's reference, then proposes a three week paid pilot that tests what the scorecard flagged: whether a two-person team answers fast enough and whether support cover is acceptable. If either fails, she talks to Harbor and Pine.
The point is the process, not the numbers: written evidence, identical questions, a weighted total, a hard rule and a pilot that tests what the paper score could not.
Pre-signing QA checklist
- Every candidate got the same brief and answered the same nine questions; two people scored independently.
- No score above 3 rests on a verbal promise, and the chosen agency scored at least 2 on ownership, data and oversight.
- I spoke to two references, one on a similar project.
- The scope lists deliverables, exclusions and acceptance tests; the pilot has a fixed fee, two to four weeks, success criteria and exit criteria.
- All accounts are in my name and I hold the admin rights; running costs are separate lines I can see.
- The data map names every tool and subcontractor, region and retention term.
- The contract names the handover pack and sign-off test, and says what I receive on termination, in what format and by when.
- A lawyer who knows my jurisdiction has reviewed the agreement.
From brief to signed pilot: a three week plan
Asset 8: Three week plan| When | What to do |
|---|---|
| Week 1, days 1 to 3 | Measure the process, write the brief (Asset 1) and fix scoring weights and scorers in writing. Step 1. |
| Week 1, days 4 to 5 | Shortlist three and send the request email (Asset 2). Step 2. |
| Week 2 | Days 1 to 3: one 45 minute call per agency, nine questions. Days 4 to 5: score independently, apply the hard rule, pick one or two finalists, send written follow-ups for any score resting on a verbal answer. Step 3. |
| Week 3, days 1 to 2 | Reference calls (Asset 5); ask for an anonymised handover pack. |
| Week 3, days 3 to 4 | Agree pilot scope, criteria and fee (Step 4); lawyer reviews against Asset 7, you create your accounts (Step 5). |
| Week 3, day 5 | Run the QA checklist, sign and set the kickoff date. Step 6 begins. |
Score us too
We said at the start that Six Figure Consulting is an agency (our AI automation agency page lists what we build). If we are on your list, send us your brief and the same email, ask the same nine questions and score us on the same sheet. A few things we will say in advance, because they are part of the scorecard: we build in the client's own accounts; our current packages and what is and is not included are listed on our packages page; our AI outbound agent is a done-with-you build that someone in-house needs to own, so it suits founders with an in-house person to manage the system better than those without; and we do not offer commission-only or pay-per-lead work. If another agency scores higher for your project, that is the right answer.
Frequently asked questions
How do I choose an AI automation agency?
Write a one-page project brief, send it to three agencies and ask each the same nine questions on scope, ownership, testing, data, human oversight, support, pricing, references and handover. Score the answers on a weighted scorecard, check two references, then prove the winner with a paid two to four week pilot.
What should I put in a project brief for an AI automation agency?
State the problem, the current process step by step with volumes and time per item, the systems involved, a success metric with a measured baseline, constraints, data sensitivity, the owner, a budget range and a timeline. The same brief to every agency makes proposals comparable.
What are the biggest red flags when hiring an AI automation agency?
Guaranteed outcomes before seeing your process, no written scope, a build that only runs in the agency's accounts, refusal to show prompts or logic, vague answers on data, and claims that no human review is needed. Each one has a follow-up question in the red flag table above.
How should an AI automation agency charge?
The four common models are a fixed project fee, a retainer, usage-based pricing and performance-linked fees. Each creates different incentives and risks. Whatever the model, keep AI usage, platform subscriptions and data tools as separate running costs that you can see and pay directly.
How long should a pilot with an AI automation agency be?
Two to four weeks, paid, on one process with real sample data, in accounts you own. Agree success criteria and exit criteria in writing before it starts, and make sure you own the build, prompts, tests and draft documentation at the end.
Do I need a lawyer to review an AI automation contract?
We recommend it. The checklist in this guide lists topics to cover, such as ownership, data processing, confidentiality, termination and liability, but it is not legal advice. Laws differ across the UAE, the US and other GCC jurisdictions, so have counsel who knows yours review the final agreement.
Sources and further reading
- Anthropic, "Building effective agents" (simplicity first, workflows versus agents, human checkpoints, stopping conditions, framework caution). Accessed 6 November 2026.
- Claude developer documentation, "Define success criteria and build evaluations" (specific and measurable criteria, task-specific tests with edge cases, automated and human grading). Accessed 6 November 2026.
- Claude developer documentation, "Data residency" (configurable inference and storage geography, zero data retention eligibility, pricing differences). Accessed 6 November 2026.
- SFC AI Outbound Agents service page (in-house owner model; commission-only and pay-per-lead not a fit).
- SFC internal onboarding, scope-of-work and build SOPs (methods adapted and anonymised; no client details used).
What to do next
Today, write the project brief using Asset 1 and start measuring your baseline. Then send the request email to three agencies and score them on the sheet. If you would like us on that list, our packages page shows how we work and what each option includes, and you can book a short call; we will tell you honestly if a different agency, a freelancer or a simple workflow suits your project better. If you are still deciding what to automate first, our guides to AI agents for lead generation and the AI appointment setter show what two common builds look like in practice.
