It depends on your segment. Observed 2026 bands put SMB engagements at $2,500 to $25,000 with $300 to $800 a month support, proofs of concept at $8,000 to $25,000, production builds at $35,000 to $150,000 and up, and mid-market programmes starting at $50,000 with a six-month minimum.
This model places a proposed engagement inside published 2026 market bands for its segment and shape, adds support, and cross-checks the total against an hourly rate range for the delivery region you picked. It models the inputs you entered; it is not a quote and it does not evaluate any specific agency.
Four different things trade under the same label, and confusing them is the reason quotes for the same brief arrive an order of magnitude apart. A proof of concept, at $8,000 to $25,000, buys an answer to a question: can this be made to work with our data, in our systems, well enough to matter. It is time-boxed, it produces a demo and a recommendation, and it should be allowed to end in no. An SMB engagement, at $2,500 to $25,000, buys a working thing for a small business: one agent, one workflow, one set of integrations, handed over and supported for $300 to $800 a month. A production build, at $35,000 to $150,000 and up, buys a system other people depend on. The difference from an SMB engagement is not features; it is evaluation, error handling, permissions, observability and the operational handover that makes it survivable. A mid-market programme, from $50,000 with a six-month minimum, buys a team rather than a project. The minimum term exists because the first two months are discovery and the value is back-loaded.
Directory-observed rates in 2026 run from $35 an hour offshore to $149 an hour for a US boutique. That is a four-fold spread on a nominally identical unit of work. The spread is real, but hourly rate is close to useless as a comparison tool on its own, because the number of hours required varies by more than the rate does. An engagement at $40 an hour that takes three times as long, requires twice the specification detail and needs a supervising engineer on your side is not cheaper. Use the implied hourly rate the way this calculator uses it: as a sanity check on a fixed price. Divide the proposal by a realistic hour estimate. If the result lands far outside the observed band in either direction, ask why. Far below usually means the scope in the proposal is not the scope in your head. Far above usually means you are buying seniority, and it is fair to ask which specific people.
Because the value in this work is back-loaded and the cost is front-loaded, and a three-month engagement reliably ends just as it becomes useful. The first month is access, discovery and finding out that the data is messier than anyone said. The second is a first working slice that surfaces every assumption that was wrong. Real compounding value tends to arrive from month three onward, once the integration surface is understood and changes stop requiring archaeology. That is the honest reason for the minimum. The less honest reason, which also exists in the market, is that minimum terms protect agency revenue predictability. Both can be true, and the way to tell them apart is to ask what happens at month three if it is not working. An agency that offers a defined checkpoint with an exit is describing the first situation. One that does not is describing the second.
Not price. Price is the least informative signal in the entire evaluation, and both ends of the range contain good and bad work. Be suspicious of a proposal with no evaluation plan. If nobody can tell you how you will know whether the thing is working, or what the acceptance criteria are, the project has no definition of done and will end in an argument. Be suspicious of a fixed price with no fixed scope. A number without a written specification is not a price, it is an opening position. Be suspicious of ownership ambiguity. Ask directly who owns the repository, the cloud accounts, the model configuration and the data, and get the answer in writing before signing. Retaining your own infrastructure is the difference between a supplier and a landlord. Be suspicious of case studies with unverifiable metrics. A 40% improvement with no baseline, no time period and no named client is a sentence, not evidence.
For a first project, almost always, because the cost of hiring is not the salary. It is the search, the mis-hire risk, the ramp, and the fact that a single in-house engineer has no one to review their work in a field where the failure modes are subtle. For a permanent capability, the arithmetic flips, and it flips faster than most agencies will volunteer. If you expect to be building continuously for two years, an in-house team is cheaper and, more importantly, retains the knowledge inside the business. The pattern that usually costs the least over three years is an agency for the first one or two builds while you learn what you actually need, with an explicit knowledge-transfer requirement written into the contract, followed by a hire once the shape of the ongoing work is clear. The build-versus-buy tool on this site scores that decision against your own timeline, budget and differentiation.
The observed 2026 SMB band is $2,500 to $25,000 for the engagement, plus $300 to $800 a month for ongoing support. The spread inside that band is driven by integration count more than anything else.
$8,000 to $25,000 in the observed 2026 band. It should be time-boxed, should end in a documented go or no-go, and should be allowed to end in no without anyone treating that as a failure.
Observed mid-market programmes start at $50,000 with a six-month minimum term. The minimum exists because value is back-loaded and the first two months are almost entirely discovery.
Directory-observed rates in 2026 run from $35 an hour offshore to $149 an hour for a US boutique. Hours required vary more than rates do, so use hourly rate as a sanity check on a fixed price rather than as a comparison method.