Ask who does the work, where client data goes, how it is accessed, what insurance the consultant carries and what exists on the last day, and check with your own E&O carrier before the first meeting. Scope the first project to one assisted workflow with a human sign-off, such as certificate requests, with a fee near a year's saving.
A consultant's pitch will talk about hours saved. Your exposure is different: a certificate, a renewal summary or an application that goes out wrong becomes your agency's error, whoever built the tool. So the hiring questions are mostly about control: where a person checks the output, what your policy says about it, and who is liable when it fails. Plan on about 11 hours of your own time before signing: two to choose the project, one to call your carrier or broker, four to interview two or three consultants, two to check references, and two to write the scope.
Pick the first project before you talk to anyone
Walk into the first consultant meeting with a candidate in hand; otherwise the consultant picks one, and the easiest project to sell is rarely the safest for you. Three options suit most small agencies. The figures below are illustrative, using a staff cost of $30 an hour.
| Option | What the AI does | Where a person checks | If it goes wrong | Illustrative yearly saving |
|---|---|---|---|---|
| Certificate requests | Reads the request email, pulls the holder and requirements, drafts the certificate from your template | A service rep compares the draft with the policy before issuing | A third party relies on a wrong statement | 60 requests a month, 6 minutes saved each: 72 hours, $2,160 |
| Renewals | Summarises the expiring policy and changes, drafts the renewal email, lists questions for the client | The account manager reviews each summary and email | Coverage misstated in a client email | 400 renewals, 15 minutes saved each: 100 hours, $3,000 |
| Intake | Collects prospect details by form, email or phone and pre-fills a submission | A producer reviews before anything goes to a carrier | A misstatement on an application | 25 enquiries a month, 12 minutes saved each: 60 hours, $1,800 |
None of the three saves a fortune in a small agency, which is exactly why you size it before the first meeting. Compare the saving with what builds cost. Contra's Zapier-freelancer hiring page lists $30 to $150 an hour, $500 to $1,500 for a simple automation across two or three apps, $2,000 to $10,000 or more for complex multi-step builds, and retainers of $1,000 to $5,000 a month. Treat those as rough marketplace figures rather than quotes. A certificate workflow that touches your inbox, your management system and one template can stay in the simple band if the scope stays tight, and then a $2,160 yearly saving covers it within a year. Scoped as a complex multi-system project, it will not.
The rule I would use to choose: the best first project is the one where a person checks every output before it leaves the agency, where the data already sits in one system, and where you can time the current process for a week. By that test, renewals and certificates beat intake. Between those two, renewals expose you less, because the output goes to a client you know rather than to a stranger who will rely on it. Certificates win on volume. If you start with certificates, build assist-only: the tool drafts and a person issues.
Three things should not be a first project: anything that answers coverage questions for clients, anything that submits to a carrier without review, and anything that handles first notice of a claim. Each puts the AI close to a statement your agency will be held to. The tutorials on automating insurance renewals with AI and AI submission intake for brokers go through the other two options in more depth, and the one on automating certificate requests without E&O risk covers the first.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowTwelve questions to put to every consultant, and what to listen for
Send the questions in writing before the meeting and compare the answers afterwards. The table pairs each question with the kind of answer worth trusting and the kind that should slow you down. The general version of this exercise is in 20 questions to ask when choosing an AI consultant; these twelve are tuned to an insurance agency.
| Question | An answer worth trusting | An answer that should slow you down |
|---|---|---|
| Who does the work, and which people, subcontractors and tools touch our data? | Named people and a list of tools with the role of each | "Our team" and no list |
| Draw the data flow for one certificate request, from email to issued document. | A diagram showing each system, each person and each check | A description of "AI handling it" |
| Where is client data processed and stored, and is any of it used for training? | Names the provider and region, says no training in writing, and offers deletion on exit | "It's all secure" |
| How will you get into our email and management system? | Named accounts, least access needed, multi-factor sign-in, removed on the last day | "Just share the admin login" |
| Where does a person check the output, and who signs off? | A named role reviews every client-facing output before it leaves | A claimed accuracy figure in place of a check |
| What insurance do you carry, at what limits, and can we see the certificates? | Technology errors and omissions and cyber cover, with certificates on request | "We've never needed it" |
| If a wrong output causes a claim, what does the contract say? | Clear indemnity and liability terms, with the cap explained | Waves the question away or caps liability at a token fee without discussion |
| What do you know about our carriers' rules for portals and AI? | Asks which carriers you use and says the agreements must be read first | "Bots log into portals all the time" |
| What exists when you leave? | A runbook, the prompts and settings, test cases, a list of every account, and staff training | "You can always call me" |
| How will we measure success? | A baseline timed on your own work before the build starts | Percentages from other clients |
| What will it cost, and where can we stop? | A paid discovery or fixed-fee pilot with an exit at each stage | A long retainer before anything works |
| Can I speak to an insurance agency you have worked with, and what went wrong? | Yes, plus an honest account of a failure | "Our clients are confidential" and nothing else |
Score the replies without looking at price. A consultant who answers the data, access and liability questions plainly has probably done this before and been caught out at least once. One who answers them with enthusiasm about how smart the model is has told you where their attention sits. The tutorial on whether agency E&O covers AI mistakes goes deeper on the coverage side, which the next two sections approach from the hiring angle.
E&O: what to ask your own carrier before the first meeting
Your errors and omissions policy decides who pays when an AI-assisted document goes wrong, and the consultant cannot answer that for you. Email your carrier or broker before you sign anything, using the same questions every time so the replies can be compared and filed. Illustrative wording:
Subject: Questions before we use an AI consultant and AI tools
1. We plan to hire an outside consultant to build AI-assisted
workflows for [certificate requests / renewals / intake]. Staff
will review every output before it leaves the agency. Does the
policy respond to errors in work produced this way?
2. Does the policy contain any exclusion, endorsement or condition
that mentions artificial intelligence, automated decisions or
third-party software?
3. Do you require us to disclose AI use at application or renewal,
and has the application form changed to ask?
4. Are there supervision conditions, such as human review of
client-facing documents, that an AI workflow must meet?
5. If a data incident comes from a consultant's tool, does this
policy respond, or is that for cyber cover?
Please confirm in writing, and tell us which answers depend on how
we describe the workflow to you.Two habits make the exercise worth doing. File the reply with the consultant's contract, because a verbal "that should be fine" disappears when a claim arrives. And describe the workflow honestly, including the human check, since a carrier that approves an imagined process has not approved yours. If the reply is vague or negative, you have learned something cheaply: narrow the project, keep a person in the loop for longer, or wait.
E&O: what the consultant's own insurance should show
Ask to see certificates for the consultant's technology errors and omissions and cyber cover, check the dates, and read the limits against the work. Then read the contract's liability clause, which matters as much as the policy. Many service contracts cap the supplier's liability at the fees paid. If you pay $3,000 for a certificate workflow and the cap is $3,000, a claim of any size comes back to your agency; the cap is a business decision you should make knowingly, not discover afterwards.
Three further contract points belong in the first draft. Indemnity: who pays if the consultant's build breaches a carrier agreement or leaks client data. Ownership: the prompts, templates and settings built for you should be yours to keep, in writing. Subcontractors: the consultant should name any, and be responsible for them. None of this needs a long contract; it needs the questions asked while you still have the choice of another supplier.
Carrier rules and portals: read before any tool logs in
Your agency agreements and the carriers' portal terms may restrict shared logins, automated access, or the use of carrier quotes and policy documents in outside tools. The wording differs by carrier, so check each one yourself and put the answers in a sheet. An illustrative page from that sheet:
| Carrier | Agreement checked | Automated or shared logins | Carrier documents in AI tools | Follow-up |
|---|---|---|---|---|
| Carrier A | Yes | Shared logins prohibited; automation not mentioned | Not stated | Asked the account rep in writing; reply due |
| Carrier B | Yes | Automated access prohibited | Not allowed without consent | No AI tool may read its quotes |
| Carrier C | Not yet | Unknown | Unknown | Check before the project starts |
The point of the sheet is the right-hand column. Any tool the consultant builds then works from a list of what each carrier allows, rather than from a general assumption that "everyone does it". If a carrier's terms are silent, ask in writing and wait for the reply. A tool that logs in first and asks afterwards is how an agency ends up explaining itself to a carrier. For the wider data-protection side, keep a written safeguards checklist for every AI tool, and ask your carrier whether it expects you to follow the insurance regulators' model AI bulletin.
Scoping the certificate pilot: a filled-in statement of work
A scope that fits on one page forces both sides to agree what "done" means. Here is an illustrative statement of work for an assist-only certificate pilot, with the numbers from the earlier table. Adapt the figures to your own timed baseline.
PILOT: certificate requests, assist-only
Baseline (timed by the agency over 5 working days): 60 requests a
month, median 12 minutes each from email to issued certificate.
In scope: reading requests that arrive in the shared inbox;
extracting holder name, address and requirements; drafting the
certificate from the agency's approved template; flagging policy
dates and endorsements that do not match the request.
Out of scope: issuing certificates; changing policy wording;
answering coverage questions; logging into any carrier portal;
using client data for anything except this pilot.
Acceptance (measured over the final 4 weeks):
- Every certificate is reviewed by a named staff member before it
is issued: 100%.
- No certificate contains wording that is not in the approved
template or the policy: 0 exceptions.
- Holder and address fields need no correction in at least 95% of
drafts.
- Median handling time, including review, at or below 6 minutes.
Access: named accounts, least access needed, multi-factor sign-in;
all removed on the last day, with written confirmation.
Stages and exit: week 1 discovery and baseline (fee A); weeks 2 to
5 build and supervised use (fixed fee B); the agency may stop after
either stage with no further charge.
Handover by the last day: runbook, prompts and settings, test
cases, list of every account and tool, one staff training session.The acceptance list is the part worth defending. "Zero exceptions" on wording turns the biggest risk into a test you can run. The 95% figure on holder and address fields makes the saving real: if staff still correct one draft in five, the minutes saved shrink. Staging the fee gives you an exit after the discovery week, when you know most about the consultant and least has been spent.
Handover documentation: what should exist on the last day
Write the handover list into the contract, with the final payment tied to it. A workable set for a small agency:
- A runbook in plain language: what the workflow does, how a request moves through it, what a person checks at each step.
- The instructions and settings the tool uses, exported in a form you can read and keep, including the certificate template and any approved wording.
- Test cases: at least ten sample requests with the expected output, including awkward ones such as a holder with two addresses or a request asking for an endorsement the policy lacks.
- An account list: every tool, login and integration the consultant set up, who owns each, and confirmation that consultant access has been removed.
- A change log showing what was adjusted during the pilot and why.
- A failure plan: what staff do when the tool is down or produces something wrong, and who to tell.
- One training session recorded for new staff, plus a named internal owner for the workflow.
The tutorial on what an AI consultant should hand over before leaving has the general version. The insurance-specific additions are the endorsement test case and the carrier sheet from earlier, which should be updated whenever a carrier changes its terms.
Pitch lines that should make you pause, and the question for each
| What you hear | Why it matters | What to ask |
|---|---|---|
| "We guarantee a percentage saving" | Nobody can know your baseline before measuring it | "Which of our own numbers produces that figure?" |
| "A fully autonomous agent for the whole agency" | No human check on client-facing output | "Show me where a person approves each output." |
| "We'll connect straight to your carriers' portals" | Carrier terms may forbid automated access | "Which carriers permit it, and where is that in writing?" |
| "Send us your policy files and we'll train a custom model" | Client data used for training is a separate decision with its own risks | "Will any client data train anything? Put the answer in the contract." |
| "You don't need to tell your E&O carrier" | Policy conditions may apply to AI-assisted work | "Will you put that advice in writing?" |
| "We need admin access on day one" | Over-broad access to client records | "Which task needs which permission, and when is each removed?" |
| "Start with a retainer and we'll find the use cases" | No scope, no acceptance test, no exit | "What is the fixed-fee first stage, and can we stop after it?" |
The rows above are the warning signs that bite hardest in an insurance setting. The question column does most of the work: say it out loud in the meeting and watch how the reply changes. A good consultant is relieved to be asked, because it gives them something concrete to price.
The certificate that promised too much
The following case is made up, but it follows a familiar pattern. A consultant built a certificate drafter that copied the wording requested in each email into the description box. One request said "include waiver of subrogation". The draft included it, although the policy carried no such endorsement. The service rep, who had been checking holders and addresses carefully, skimmed the description field because the tool had been right for weeks. The certificate went out, and the third party relied on it.
The workflow was not at fault so much as its scope. The fix was to stop free-form wording altogether: the tool now fills only fields from the approved template, and anything the request asks for that the template or policy lacks is listed in a separate "needs a decision" box instead of being written into the certificate. The acceptance test in the statement of work, zero wording not in the template or policy, would have caught it in week two. That is the case for writing the test before the build starts, and for having a person who owns the check rather than a tool that is usually right.
Four weeks from first call to signed scope
- Week 1, about 3 hours: choose the candidate project using the table above, time the current process for five working days, and send the carrier questions.
- Week 2, about 4 hours: interview two or three consultants with the twelve questions, asking each for the data flow and the insurance certificates.
- Week 3, about 2 hours: check references, speaking to at least one insurance agency, and compare the carrier's reply with what each consultant said about E&O.
- Week 3 or 4, about 2 hours: write the one-page scope with acceptance tests and stage gates, then have your broker or lawyer read the liability and data clauses.
- Week 4: sign the discovery stage only. Release the build stage when the baseline is measured and the acceptance numbers are agreed.
That is the 11 hours promised at the start, spread over a month. The hours cost less than most first retainers, and they are what lets you tell, six weeks later, whether the consultant delivered what the scope said.
Questions agency owners ask before hiring an AI consultant
Should we tell our E&O carrier before hiring an AI consultant?
Ask before the project starts, not after something goes wrong. Put the questions in writing, covering exclusions or conditions that mention AI, supervision requirements and whether the application now asks about AI use, and keep the reply. Policy wording varies, so rely on your carrier's written answer rather than on what the consultant says is normal.
Is a fixed-fee pilot better than an hourly engagement?
For a first project, usually yes, because it fixes the cost and forces a written scope with acceptance tests. Hourly work suits discovery, when nobody yet knows what exists. A sensible pattern is a short paid discovery, then a fixed fee for one workflow, with an exit at each gate. Avoid retainers until a pilot has run through a full renewal cycle.
Can the consultant log into carrier portals for us?
Not until you have read each carrier's agency agreement and portal terms, which may restrict automated or shared logins. Even where it is allowed, use named accounts with the least access needed and multi-factor sign-in, never shared credentials. If a carrier's terms are unclear, ask the carrier in writing; do not let a tool log in first and ask afterwards.
What if the consultant wants to use our client data to improve its own tools?
Refuse unless you have agreed to it in writing and checked your carrier and client terms. Client files you hold are covered by duties to your clients and often by carrier agreements. The contract should say client data is used only for your project, is not used to train any model, and is deleted or returned at the end, with written confirmation.
Further reads
- AI Insurance Agency Management System: What to Ask Before You Switch — What to ask before switching your management system for AI
- GLBA Safeguards for Insurance Agencies: An AI Tool Checklist — A safeguards checklist for any AI tool touching client data
- NAIC AI Model Bulletin Explained for Independent Insurance Agencies — The regulators' model AI bulletin, explained for agencies
- AI Consultant Red Flags: 12 Warning Signs to Walk Away From — Twelve general warning signs to walk away from
- Who Owns the AI Workflows a Consultant Builds for You? — Ownership of what the consultant builds for you
- Should You Sign an NDA Before Sharing Data With an AI Consultant? — Whether and how to sign an NDA before sharing data
- OpenAI Consulting Firm vs Independent Consultant for a Small Business — A ten-question test for DIY, solo or firm, published fee bands, who owns your prompts and workflows, and a ChatGPT setup scope you can hand to either.
- AI Cross-Selling for Insurance Agencies: Account Review Prep, Not Advice — AI prepares the file summary and the questions; the licensed agent advises. A prep routine, prompts and a sign-off log for account reviews.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Contra's Zapier-freelancer hiring page for market rates (checked November 2026). Illustrative figures elsewhere are labelled as such; E&O and carrier terms are for your own policy and carriers to confirm.