AI Agent Development Companies: What to Ask and What a Quote Covers

Before agreeing a quote with an AI agent development company, ask what exactly the agent will and will not do, what data and system access it needs, how it will be tested against your real cases, who owns the prompts and code, and what it costs each month to run once the build is finished.

A sound quote itemises those five things in plain words. A weak one gives a single price for "an AI agent" and leaves scope, ownership and running costs to be discovered later. A ranking of agencies cannot tell you which of them will handle those five things well for your job; the answers to the questions below can. Fee figures are left out on purpose, because the right number depends on your scope, not on a published rate.

Follow me on Instagram@sagnikteaches

Start by writing the job in one paragraph

You cannot judge a quote for an agent until you have written down what you want one to do. Two or three sentences are enough. An agent in this sense is software that takes a task, decides which steps to take using a language model, and uses tools such as your email, calendar or booking system to carry them out. It differs from a chatbot, which only answers.

Connect on LinkedInSagnik Bhattacharya

Say a removals firm wants an agent to deal with quote requests. The scope paragraph: "When a customer emails or fills in the form, the agent reads their message, asks for anything missing (move date, addresses, rooms, stairs, access), calculates an indicative quote from our rate sheet, drafts a reply for a coordinator to approve, and creates a record in our booking system. It never sends a final price or confirms a date without human approval."

Subscribe on YouTube@sagnikteaches

That paragraph does three jobs. It gives the company something concrete to price, it fixes a human-approval point, and it gives you a yardstick for every question below. If you want help turning a rough wish into a proper brief first, how to brief a developer on a custom AI workflow covers it. The questions below are the ones specific to agents, which act in your systems rather than only answering.

Scope: five questions that fix what you are buying

1. "List what the agent will do, step by step, and what it will not do." A good answer mirrors your paragraph and adds boundaries: the agent does not issue refunds, does not edit customer records without approval, does not answer outside the topics you name.
2. "What happens when it is unsure?" Look for a defined handoff: the case goes to a named person, with the conversation and the agent's reasoning attached. If the answer is "it will usually figure it out", press for detail.
3. "Which systems will it connect to, and which actions can it take in each?" Read access and write access are different risks. An agent that can read the calendar is safer than one that can also delete events.
4. "What is in version one, and what is deferred?" Agents invite scope growth. Ask for a written list of what is out of scope, so a later "can it also do X?" is a priced change, not an argument.
5. "Who does the work, and who is accountable?" Find out whether the people on the sales call will build it, whether any work is subcontracted, and who your single contact is.

Here is how good and weak answers differ on the second question, illustrative:

WEAK:
"The agent is very capable and handles edge cases intelligently."

SOUND:
"If the customer mentions a piano, a flight of more than two stairs, or an
address outside the rate sheet's areas, the agent stops, tells the customer
a coordinator will follow up, and posts the case to the coordinators'
shared inbox with the full thread and the fields it could not fill."

The sound answer names triggers and a destination. The weak one makes a claim you cannot test.

Data and access: what the company needs and what you keep

An agent is only as safe as the access you give it. Ask these questions before anyone sees a login.

• "Exactly what data will the agent see?" Ask for a table of data types, where each comes from, and why it is needed. Customer names, addresses and payment details deserve a line each.
• "Where is the data processed and stored, and does the model provider keep it?" The answer should name the model provider and the plan or agreement in use, and say whether your data is used for training. Business and API terms differ from consumer terms, so ask for the specific terms, and read them.
• "What credentials will you need, and how are they held?" The answer should be a dedicated account or token with the narrowest permissions, stored in a secrets manager, not shared logins sent by email.
• "How can the agent be made to misbehave, and what have you done about it?" The honest answer includes prompt injection, where text inside an email or web page tries to give the agent new instructions. Ask what the agent can do if that happens. What prompt injection is and whether a small business should worry explains why this matters for any agent that reads outside text.
• "Can we run it on a test copy first?" A sandbox or test mailbox with dummy customers is the right place to try everything before live data is connected.
• "How do we revoke access, and what happens to our data at the end?" You want a clear exit: tokens revoked, data returned or deleted, in writing and with a timescale.

If personal data about customers is involved, including customers in the EU, ask the company which data-protection laws it works to and whether it signs a data-processing agreement. Do not accept "we are compliant" as an answer without seeing the paper.

Testing: how you will know it works before you pay in full

Many agent projects fail at handover, not at build. The demo worked on the easy cases, and the first real customer found the hole. Ask for testing to be part of the quote and the milestones.

1. "Will you build a test set with us before you build the agent?" The answer should be yes: 30 to 50 real, anonymised examples, each with the output you would accept, including awkward ones such as missing information, rude messages, and requests the agent should refuse.
2. "What score counts as passing?" Agree it in writing. For the removals firm: all 50 cases handled without a wrong price sent, at least 45 handled without human correction, and every refusal case routed to a person.
3. "Will you show me the failures?" A company that reports only successes is hiding something. Ask for the failing cases and what was done about each.
4. "Can I run the tests myself at handover?" You should be able to run the same set after delivery, and again whenever the model, a prompt or a connected system changes.
5. "What is the plan for a live pilot?" A sound answer is a few weeks with a human approving every output, then a gradual removal of approval for the categories that prove reliable.

For a longer treatment of the testing step in a customer-facing context, how to test a customer chatbot before it goes live applies almost line for line.

Ownership: prompts, code, accounts and the exit

The agent is a bundle: prompts and instructions, code or workflow definitions, accounts with model and tool providers, test sets, and documentation. Each has an owner, and the quote should say who. Put these questions in writing:

• Prompts and instructions. Do we own the final text? Will we receive it in a readable file, not just see it in a dashboard?
• Code and workflows. Is everything bespoke assigned to us on payment? Which components are the company's own reusable tools, and on what licence do we use them after the project?
• Accounts. Are the model, automation platform and hosting accounts in our name, with the company as an invited user, or in theirs? Accounts in the supplier's name are the commonest form of lock-in.
• Test sets and logs. Do we own the tests and the conversation logs?
• Documentation. Will there be a plain-English runbook covering how it works, how to pause it, who to call, and how to change a rate or a rule?
• Third-party terms. Does any component carry restrictions, such as non-commercial licences or usage caps?

Moving to another supplier or taking the agent in-house should be possible without a rebuild. How to keep your data and prompts portable covers the practical safeguards, including how to export what you need regularly.

Running costs: the part the build quote usually leaves out

The build price is one number. The cost of keeping the agent alive is several, and most of them recur. Ask for each of these to be written down, even if the answer is "we will estimate after the pilot". No amounts appear here because they depend on supplier and volume; the point is to make sure every line exists in the quote.

Model usage deserves a note. Providers publish per-token prices that change (Anthropic's API price list, for example, is set per million input and output tokens, with output tokens priced higher), and a long conversation costs more than a short one. Ask the company to run a usage estimate from your test set: the number of model calls and tokens it used, scaled to your expected volume. If they cannot, they have not measured it. How much a custom AI assistant costs a small business sets out the cost categories in more detail.

Model changes deserve a line of their own in the support terms. Providers retire older models on a published schedule: Anthropic's deprecation page, for instance, promises at least 60 days' notice before a publicly released model is retired, after which requests to it fail. An agent built on a model that is retired mid-contract needs re-testing against your full test set on the replacement, and the quote should say who does that work and who pays for it.

A filled-in quote outline for the removals agent

To see how the pieces fit, here is how a complete quote for the removals firm's agent might be laid out. No amounts appear because they depend on the supplier; the structure is what to look for.

Payments attached to phases 2 and 4, not to calendar dates, give the company a reason to finish properly. If a supplier insists on payment before acceptance, ask why; the usual reason is cash flow, which can be met with a smaller deposit rather than waiving the tests.

The first month after launch

The quote should also cover what happens once the agent is live, because that is when real customers find the cases nobody imagined. Ask for a weekly review in the first month: the company and your owner read a sample of real conversations, log each correction, and decide whether it needs a prompt change, a new rule or a new test case. Every real failure should become a test case, so the set grows with experience.

Agree in advance who can pause the agent. The answer should be you, immediately, with a documented switch, not a support ticket. Also agree what triggers a pause: a wrong price sent, a message to the wrong customer, an unexpected spike in model usage, or a connected system changing its format. The agent is software that acts for you, so the owner needs the same control a manager has over a new employee in their first weeks.

Reading the quote itself: what a complete one itemises

A quote you can sign has sections, not a single price. Check for each of these:

1. The scope paragraph, and a list of exclusions.
2. Phases with deliverables: discovery, build, test, pilot, handover.
3. Acceptance criteria for each phase, tied to the test set.
4. Payment milestones attached to acceptance, not to dates.
5. Your responsibilities: who supplies examples, access and approvals, and by when.
6. Change procedure: how a new request is estimated and approved.
7. Ownership and licence terms, as above.
8. Running-cost estimate and support terms after launch.
9. Exit terms: access revocation, data return, handover documents.

Compare this list with a typical thin quote, illustrative: "AI agent for customer enquiries: design, development, deployment and 30 days of support." That sentence answers none of the questions. It does not say what the agent does, how it is tested, who owns it, or what happens on day 31.

Using AI to check a quote against this list

Once you have a quote, you can use an assistant to find gaps, as long as you treat the answer as an index and not a verdict. Remove names and prices if you prefer, then paste the quote text.

Below is a quote from a company proposing to build an AI agent for my
business. Compare it with this checklist and for each item say PRESENT,
PARTLY or MISSING, quoting the sentence that supports your answer:
scope and exclusions; human handoff; systems and access level; data
handling and model provider terms; test set and pass mark; pilot;
ownership of prompts, code, accounts and logs; running costs; change
procedure; support terms; exit. Then list the five questions I should send
back to the company. Do not judge whether the price is fair.

QUOTE:
[paste text]

Illustrative output on the thin quote above:

Scope and exclusions: MISSING ("design, development, deployment" names
  no tasks).
Test set and pass mark: MISSING.
Ownership: MISSING.
Running costs: MISSING.
Support: PARTLY ("30 days of support", no response time or scope).
Questions to send: 1) What will the agent do, and what will it refuse?
2) How will you test it, with whose examples? ...

What you would fix: ask the model to quote the sentence for every PRESENT, then check a few yourself. It will sometimes mark something PRESENT because a related word appears, and your own read of the clause matters more than its tick.

Red flags that should slow you down

• A price before any questions. A company that quotes without asking about your process is pricing a product it already has, not your problem.
• Promised accuracy figures with no test method. "95 per cent accuracy" means nothing until you know what was measured on whose data.
• Everything is "proprietary". You may use a platform you cannot export from, which is lock-in.
• Accounts in their name. Your model, automation and hosting accounts should be yours.
• No human approval step for anything that sends money, confirms a booking or emails a customer, at least during a pilot.
• Vague answers about the model. You should be told which model provider is used and on what terms, even if the company changes it later.
• Pressure to sign this week because of a "slot" or "discount".
• No references you can phone. Case studies on a website are marketing; a customer who will talk to you is evidence.
• Silence about failure. Ask what went wrong on their last project. A good company has a story and a lesson.

When a simpler route beats an agent

Not every job needs an agent. If the steps never vary, a rules-based automation is cheaper and easier to test, and if the job is only drafting text, a prompt in a shared Project may do. Before paying for custom work, read whether to build or buy and ask the company to explain why a simpler tool would not do. A company that argues itself out of work is one you can trust with the rest. If you would like someone to go through a quote with you, that is what my AI implementation consultation covers, and you can reach me here.

Questions owners ask when a quote lands

Should I choose a fixed fee or time and materials?

A fixed fee suits a tightly written scope with clear acceptance tests. Time and materials suits discovery or work whose shape will change. Either can go wrong: fixed fees invite scope disputes, and open-ended hours invite drift. Whichever you pick, ask for a cap, milestones and a written definition of done.

Who should own the prompts and code?

You should, for anything built specifically for your business, written into the agreement in plain words. The company may keep its generic tools and libraries. Ask for a list of what is bespoke, what is reused, and any licence terms on the reused parts, so you can move to another supplier without rebuilding.

How do I know whether the agent actually works?

Agree a test set before the build: 30 to 50 real examples with the answer you want for each, including awkward ones. The company runs them, shows you the results and the failures, and you re-run them yourself at handover. A demo that only shows the easy cases proves very little.

Do I need a technical person on my side?

You need an owner who understands the business process and can approve test results, not a developer. That person answers questions, supplies examples, signs off acceptance and monitors the agent afterwards. If nobody on your team can do that, factor it into the plan before you sign.

Further reads

• How to Choose an AI Consultant: 20 Questions to Ask First — Twenty questions for choosing an adviser before a developer.
• AI Consultant vs Agency vs Freelancer: Which Should You Hire? — Which kind of supplier suits the size of job.
• How to Prepare Your Small Business for AI Agents — What to have ready before any agent project starts.
• Questions to Ask an AI Vendor Before You Sign Anything — A wider list for any AI product vendor.
• How to Evaluate an AI Implementation Proposal or Quote — The general method for judging any AI proposal or quote.
• AI Automation Consulting for Small Business: Scope, Cost and Fit — What a first AI automation engagement covers, how day rates and fixed fees compare, a payback test, and questions to ask before you hire.
• AI Automation Setup for Small Businesses: DIY or Hire It Done? — A decision rule for DIY versus hired AI automation setup, built on hours saved, data sensitivity and who on your team can maintain it.
• AI Chatbot Development Company, Freelancer or Platform? A Small-Firm Choice — Compare a chatbot development company, a freelancer and a chatbot platform on scope, ownership of the bot, ongoing fees and contract terms.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic API pricing page (per-million-token billing) and model deprecations page (at least 60 days' notice before retirement); general contracting practice. No fee figures are quoted because rates vary by supplier and scope. The worked scope and the sample quote are illustrative.