How much does AI consulting cost? It depends on things you can name before the first call.

How do AI consultants price their work?

There are five common models, and many firms combine two or three of them depending on the phase. None is better in general. Each one moves risk between you and the consultant in a different direction, and the right one depends on how well the work can be described before it starts.

A pattern that tends to work: a short, bounded engagement first to find where the work is, then a fixed fee for anything that needs building, then optional ongoing help. Each step earns the next.

What drives the price up or down?

Six things account for most of the difference between a small quote and a large one. You can answer most of them before you talk to anyone, which makes every quote you get easier to compare.

• Team size. Training four people who own one process is a different job from rolling a tool out to forty. More people means more sessions, more permissions and more follow-up.
• Number of workflows. One repeated process is a contained problem. Five processes across three departments is five problems, and they rarely share a solution.
• Systems to integrate. Work that stays inside a chat window or a folder of files is cheap. Each live system the AI must read from or write to (a CRM, an accounting package, a property management platform) adds authentication, permissions, testing and a way for it to break.
• Data readiness. If the inputs are consistent files in one place, the work goes quickly. If they are scattered across inboxes, scanned PDFs and a spreadsheet only one person understands, cleanup can be the largest line item.
• Security review. Regulated industries, client confidentiality obligations or an IT team with a vendor questionnaire all add time. That time is worth paying for, but it belongs in the estimate.
• Who maintains it. A prompt library your team runs by hand needs almost no upkeep. A deployed agent that calls live systems needs monitoring, updates when an upstream API changes, and a named owner.

What costs sit outside the consultant's fee?

Model usage, software seats and hosting are usually billed by the vendors directly, not by the consultant. At pilot volume they are typically a small monthly amount. They grow with usage, so ask for an estimate at the volume you expect in a year, not only at launch.

• Seats. Per-user subscriptions to an assistant such as Claude or ChatGPT on a business plan. For many small teams this is the largest recurring line.
• Model or API usage. Automations and agents that call a model through an API are metered by usage. A handful of runs a day costs little. Thousands of documents a day is a real budget line.
• Hosting. A deployed agent or connector runs somewhere, such as a serverless platform or a cloud account. At small scale this is often cheap; it is still a bill with your name on it.
• Automation platforms and data. Tools such as Zapier or Make meter usage on their own plans, and licensed data sources have their own terms, some of which restrict AI use entirely.

A good proposal lists every one of these, says who pays each vendor, and puts the accounts in your name so nothing is stranded if the engagement ends.

What are the red flags in an AI consulting quote?

The warning signs are mostly about scope and ownership, not the size of the number.

• A price before anyone has looked at the workflow, the files or the systems.
• A scope written in outcomes ("transform operations") rather than deliverables you could check.
• No plan for measuring whether it worked, or a baseline nobody recorded.
• Model and hosting costs marked up and resold, with the accounts in the consultant's name.
• A long retainer as the only way in, before anything has been built or proven.
• Silence on who owns the prompts, code and configuration when the work is done.

How does altr structure an engagement?

We do not publish fees. We quote on a call and put the number in writing before anything is committed, because the right figure depends on the team's size, location and how far along it already is. What we can publish is the shape.

• Enablement first. Hands-on work against the team's own tools, as a working session for the one to four people who own a process, or a team workshop for up to ten. The output is prompts and workflows the team runs by hand. See enablement.
• A build, only if enablement finds something. One fixed fee against written scope, with hours named per workstream. See engineering.
• A retainer, only after enablement or a build. Up to ten advisory hours a month, month to month or on a three or six month term, with unused hours rolling one month.
• Vendor costs billed by the vendors. Model and infrastructure costs go directly to the client and are a small monthly amount at pilot volume.

The whole method is on how we altr work.

What should you ask before signing?

Ask questions that turn the proposal into something you can check later.

1. What exactly is delivered, and how will we know it is finished?
2. Which pricing model is this, and what happens if the work runs over?
3. What will the vendor costs be at pilot volume and at a year's expected volume?
4. Whose name are the accounts, API keys and hosting in?
5. Who owns the prompts, code and configuration at the end?
6. What does maintenance look like, and who does it after you leave?
7. How will we measure whether this saved time or money?

For the wider checklist, see how to choose an AI consultant.