The AI consulting business: four offers that sell, and how to price them
An AI consulting business sells four things: a fixed-fee audit, a scoped pilot, an implementation retainer and a fractional AI officer. Each has a clear buyer, a price you can build from hours and risk, and a defined end. Agencies that add 'AI' to the homepage without one of these offers sell nothing, because the buyer cannot tell what they would be paying for.
Most agencies I know have added a line about AI to their homepage. Very few have sold an AI engagement at a decent margin. The gap has a simple cause: “we do AI” is a capability, and clients buy offers. An offer has a buyer, a deliverable, a price and an end.
I have sold all four of the offers below from inside a software agency. This page covers what each one is, how I build the price, and where the delivery goes wrong. It sits within the AI hub, and the reason a services firm needs a line of work like this is set out in is software engineering dead.
Why “AI services” on the homepage sells nothing
A buyer at a mid-sized company has a vague mandate from their board to “do something with AI” and a specific fear of wasting money. They land on a page that says the agency offers AI strategy, AI development and AI integration. They learn nothing about what happens first, what it costs, or what they will have at the end. So they do nothing, or they hire whoever gave them a concrete first step.
Three more mistakes follow from the same root:
- Selling technology in place of a result. Clients have a slow quoting process or a support backlog. That is the thing they will pay to fix.
- Leading with a large build. Nobody signs a six-figure project for an unproven idea from a vendor they have not worked with.
- Free discovery. Weeks of unpaid workshops, after which the client takes the ideas in-house.
The fix is a ladder of offers. Each step is small enough to buy, pays for itself, and earns the right to propose the next.
Offer 1: the audit
What it is. A two-week, fixed-fee assessment. You interview the people who do the work, look at the data and the systems, and deliver a ranked list of opportunities with a rough cost, a rough benefit and the main risk for each. You finish with one recommended pilot, scoped and priced.
Who buys it. An operations director, a COO, or a founder with a mandate and no plan.
Pricing it.
| Item | Hours | Rate | Amount |
|---|---|---|---|
| Senior consultant: interviews, data review, write-up | 45 | $220 | $9,900 |
| Partner: kickoff, review, final presentation | 10 | $220 | $2,200 |
| Total | 55 | $12,100 |
I quote $12,500 fixed, half on signature. If the client signs the recommended pilot within 30 days, I credit half the audit fee against it. The credit gives the client a reason to decide, and it costs less than a second sales cycle.
Where it goes wrong. The audit turns into a strategy deck with twenty ideas and no ranking. A good audit says: do this one first, here is why, here is what it costs, and here are the three you should ignore.
Offer 2: the pilot
What it is. A four to eight week build that proves one use case on the client’s real data, with a measurable pass mark agreed before work begins. For example: the system extracts the eight required fields from supplier invoices and matches a person’s result on at least 95 of every 100 documents in an agreed test set.
Who buys it. The audit client, now with a specific business case.
Pricing it. Take a six-week pilot.
| Item | Hours | Rate | Amount |
|---|---|---|---|
| Senior engineer: build, evaluation, integration | 180 | $200 | $36,000 |
| Lead: scoping, client sessions, review | 40 | $200 | $8,000 |
| Subtotal | 220 | $44,000 | |
| Contingency at 20% | $8,800 | ||
| Total | $52,800 |
I quote $52,500 fixed, in three payments. Model usage and hosting run on the client’s own accounts, at their cost, from the first day. That keeps a variable cost you cannot control out of your fixed price, and it means the client owns the infrastructure at handover.
The contingency is higher than I use on ordinary build work. The reason is the data. You find out how bad it is in week two.
Where it goes wrong. No pass mark. Without an agreed test set and a threshold, the pilot never ends, because someone can always find an example where the output is poor. Write the acceptance test into the statement of work. This is the same discipline that stops scope creep on any project, applied to a system whose output varies.
Offer 3: the implementation retainer
What it is. After a successful pilot, the system has to be rolled out, watched and extended. Models change, data drifts, users find edge cases, and the client wants the next use case. A retainer covers a fixed amount of senior capacity plus monitoring.
Who buys it. A client with a working pilot and no internal team to own it.
Pricing it.
| Item | Monthly amount |
|---|---|
| 50 hours of senior engineering capacity at $200 | $10,000 |
| Monitoring, evaluation runs and incident response | $2,500 |
| Total | $12,500 |
Six-month minimum, billed in advance. Unused hours do not roll over beyond one month. The mechanics, including overage rates and when a client should move back to a project, are in retainer pricing.
Where it goes wrong. The retainer has no backlog, so the client feels they are paying for nothing. Keep a visible, prioritized list of improvements and report monthly on accuracy, volume and time saved.
Offer 4: the fractional AI officer
What it is. Senior advice on a monthly fee, with no delivery attached. You sit in on leadership meetings, set priorities, review what vendors propose, write the usage policy, and stop the company from buying things it does not need.
Who buys it. A company with 50 to 500 staff that needs someone accountable for AI decisions and cannot justify a full-time executive.
Pricing it. Three days a month at $2,500 a day is $7,500 a month, on a three-month rolling term. It has the highest margin of the four because it is one senior person’s time with no delivery risk. It also depends entirely on that person, so it does not scale past what your most senior people can carry.
Where it goes wrong. Conflict of interest. If you advise on what to build and then bid to build it, say so openly and let the client take competing quotes. Clients forgive a declared interest. A hidden one ends the relationship when it surfaces.
How the four fit together
One client taking the first three steps in a year looks like this:
| Engagement | Revenue |
|---|---|
| Audit | $12,500 |
| Pilot, less the $6,250 audit credit | $46,250 |
| Retainer, nine months at $12,500 | $112,500 |
| Year-one total | $171,250 |
Most of the money is in the retainer. The audit and the pilot are how you earn it. An agency that tries to skip to the retainer has nothing to point at when the client asks why they should trust it.
For how to set rates in the first place, see how to price consulting services. Once an offer is stable and repeatable, it can be packaged with a published price, which I cover in productized services.
Delivery risk, and how to contain it
AI work fails in ways ordinary software work does not. Price and contract for each.
Output varies. The same input can produce different output. Define quality as a measured rate on a test set. Never promise that the system will always be right.
The data is worse than described. Put a data review in week one of the pilot, with a clause allowing a re-scope if the data does not meet stated assumptions.
Running costs surprise the client. Estimate the monthly usage cost at expected volume during the pilot and put the figure in the handover report.
People do not adopt it. A system nobody uses is a failed project even if it passes its test. Budget time for training and for changing the process around the tool.
Wrong output causes real harm. For any decision with money, legal or safety consequences, keep a person in the loop and say so in the contract. Cap your liability. Have the client accept responsibility for how outputs are used.
The underlying models change. A provider updates or retires a model and behavior shifts. The retainer exists partly for this. Keep the test set so you can rerun it after every change.
Staffing it
You need fewer specialists than you think. The core skills are careful engineering, evaluation, data handling and the ability to understand a client’s operations. A strong senior engineer with real curiosity learns the rest on two or three projects.
The harder problem is the first few engagements, when you have sold something your team has not delivered before. You can hire ahead of revenue, train on a client’s time, or bring in a specialist partner for the first projects while your own people learn. I lean toward the third, and the reasoning is in when to partner instead of hire.
If you are an individual thinking about this as a solo practice, the route in is covered in how to become an AI consultant. For the delivery role itself and how buyers evaluate it, see what an AI implementation consultant does.
Where to start on Monday
Write the audit offer on one page: who it is for, what they get, how long it takes, what it costs. Send it to five existing clients who have mentioned AI in the last six months. Existing clients already trust you with their systems, which is the hardest thing for a new AI vendor to earn. One of the five will say yes, and that audit becomes the proof for the next ten conversations.
Common questions
- What does an AI consulting business actually sell?
- It sells help putting AI to work inside a client's operations or product. In practice that means four offers: an audit that finds and ranks the opportunities, a pilot that proves one of them on real data, a retainer that runs and extends what was built, and ongoing senior advice for leadership.
- How much should I charge for an AI audit?
- Build it from the hours and your target rate. A two-week audit with 55 hours of senior time at a target of $220 an hour comes to about $12,000, so a fixed fee near $12,500 is reasonable. Charging nothing for the audit attracts clients who will not pay for the pilot either.
- Is AI consulting profitable for a software agency?
- It can be more profitable than build work because clients pay for judgment and outcomes, and the engagements chain together. The margin depends on fixed scoping, clear acceptance criteria and a client who owns their own running costs. Open-ended experiments on an hourly basis lose money and goodwill.
- What is the biggest risk in delivering AI projects?
- Promising a level of accuracy before you have seen the client's data. Model output varies, and data in real businesses is messier than anyone admits in the sales call. Agree a test set and a measurable threshold in the pilot, and keep a person in the loop for decisions that carry real cost.
- Do I need machine learning researchers to run an AI consultancy?
- For most client work, no. The work is integration, data handling, evaluation and process design built on existing models. You need strong engineers who are careful about testing and someone who can understand a client's operations. Research skills matter only for a narrow set of specialist problems.