How much AI development costs in 2026
Build cost by project type, running cost, engagement models, and how to tell whether a quote is honest.
AI development projects start at $4,500 and most cost $6,000 to $26,000 to build with a specialist firm like MacroCoderz, and $60 to $3,000 a month to run. A single-purpose chatbot sits at the bottom of that range; a multi-tenant RAG system or agent platform at the top. US agencies charge 3-5x more for the same engineering.
The number is set by integration count, data quality and evaluation rigour, not by which model you use. Every figure below is labelled as either MacroCoderz's published rate or a market-wide range, so you can tell which is which.
How much does AI development cost by project type?
These are our own published ranges, the same ones on the pricing page. The low end assumes clean data and one integration; the high end assumes messy data, several integrations and a compliance review. The same brief can differ by a factor of three depending on how many systems it has to touch.
| Project type | Build cost (MacroCoderz) | Timeline |
|---|---|---|
| AI readiness audit, written | $2,000-3,500 | 1-2 weeks |
| Evaluation harness for an existing system | $3,000-6,000 | 2-3 weeks |
| Single-purpose chatbot on your own content | $4,500-6,000 | 2-3 weeks |
| AI chatbot + knowledge base with handoff | $4,500-5,500 | 3-4 weeks |
| Workflow automation suite | $4,500-7,500 | 4-6 weeks |
| Document search / RAG over one corpus | $5,500-9,000 | 5-7 weeks |
| Multi-tenant RAG with isolation and SSO | $12,000-20,000 | 8-12 weeks |
| Custom agent platform | $14,000-26,000 | 10-16 weeks |
AI agents specifically
Agents are priced in three tiers. The full breakdown, including running cost at different volumes, is in the guide to what AI agents cost, and the service itself is AI agent development.
| Agent tier | Scope | Build cost | Timeline |
|---|---|---|---|
| Single-purpose agent | One job over your own content, one or two integrations | $4,500-8,000 | 2-4 weeks |
| Production agent | Multi-step, a few integrations, human handoff and approval paths | $8,000-18,000 | 4-7 weeks |
| Agent platform | Several agents or systems, shared tooling, permissions and evaluation | $14,000-26,000 | 10-16 weeks |
Retrieval, integration and products
Retrieval over your own documents starts at $5,500 for one corpus and runs $12,000-20,000 for a multi-tenant system with isolation and SSO - see RAG systems. Adding AI to a product you already have starts at $4,500, and most integrations land between $6,000 and $18,000. A launchable first version of a new AI product starts at $4,500 and most land between $8,000 and $25,000.
What drives the cost of an AI project up or down?
In rough order of how much they move the total. Model choice is not on the list until the end, because it is almost never the expensive decision.
1. Integration count
Each external system is an auth flow, a rate limit, a data model to learn, a sandbox to get access to and a failure mode to handle. One integration is about a week. Five is not five weeks - it is more, because they interact. This is the single biggest driver and the one most estimates miss.
2. Data quality
Clean, consistently structured documents in one format are cheap. Scanned PDFs, six naming conventions and three systems of record that disagree can be half the project. It is the most common reason an estimate doubles, and it is knowable in week one if someone looks at ten real files.
3. Evaluation rigour
A demo needs no evaluation. A system people rely on needs a labelled test set, regression runs on every prompt change and agreed thresholds for shipping. That is typically 15-25% of a build, and it decides whether the system still works in month six.
4. Compliance burden
GDPR scoping, data residency, audit logging, per-tenant isolation and a DPA are real engineering, not paperwork. HIPAA more so. Budget 20-40% on top of the equivalent unregulated build.
5. How wrong it is allowed to be
A system that drafts something a human sends is cheap. One that sends it needs confidence thresholds, an approval path, an audit trail and a rollback. The difference between "suggests" and "acts" is often 2x.
6. Latency and volume
Under 3 seconds is usually free; under 1 second means caching, smaller models on the hot path and a lot of measurement. Cost scales with tokens and documents, not users - ten thousand documents and ten million behave differently at the embedding, storage and re-ranking layers.
What does AI cost to run after launch?
Running cost is a separate budget line from build cost, and it scales with success: the more a system is used, the more it costs. We bill none of it ourselves. Model API spend and cloud hosting go to your own accounts at cost, so you keep the vendor relationships when the engagement ends.
| System | Typical running cost / month | What it covers |
|---|---|---|
| Workflow automation | $60-250 | Model calls only where judgement is needed, plus the automation platform |
| RAG over one corpus | $150-600 | Embeddings, vector storage and inference |
| Production AI agent | $200-3,000 | Model APIs, vector storage and hosting |
- Model API spend (OpenAI, Anthropic, Cohere and similar): a typical production RAG system runs $150-800 a month.
- Cloud infrastructure - hosting, vector database, storage: $50-400 a month for most builds, more at scale.
- Third-party licences and SaaS seats the system depends on.
- Maintenance. Models are deprecated, APIs change and documents move; budget a few days a quarter, or keep an engineer on it.
The largest lever on running cost is caching repeated queries, followed by routing easy requests to a small model with a larger one as fallback - which typically cuts inference cost 60-75%. Both are architecture decisions made during the build, which is why running cost belongs in the scoping conversation, not after launch.
How does cost change by engagement model?
The same system costs very different amounts depending on who builds it and how they bill. The rates below are market-wide ranges except for the MacroCoderz row, which is our published rate.
| Engagement model | Typical rate | Best for | Trade-off |
|---|---|---|---|
| US agency (market) | $150-300/hr | On-site enterprise work, procurement that requires a domestic vendor | 3-5x the cost for the same engineering |
| EU agency (market) | EUR 120-250/hr | EU data residency and an EU legal entity on the contract | Similar cost premium, narrower AI bench |
| Freelancer (market) | $60-150/hr | Scoped tasks under $5,000, or one skill for two weeks | No bench, no continuity, no cover |
| In-house hire (market) | $180-260k/yr loaded | Long-term ownership of an AI roadmap | 3-6 months to fill, 4-8 weeks to ramp |
| MacroCoderz fixed project | From $4,500 | A defined outcome with a deadline | Fixed scope - changes are quoted in writing |
| MacroCoderz dedicated engineer | $2,800/engineer/month | An ongoing roadmap without a hire | No on-site presence |
Fixed project or dedicated engineer?
Choose a fixed-price project when you can write down the outcome and the deadline: you pay for a delivered system, not for hours. Choose a dedicated AI engineer when the work is a roadmap that keeps changing. The trade-offs are set out side by side in staff augmentation vs a project build.
The year-one arithmetic for one engineer
Comparing a salary with a monthly rate understates the in-house cost. For one mid-to-senior AI engineer, year one looks roughly like this: a US hire at $180,000 base comes to about $254,000 once payroll tax, benefits and a recruiting fee are added; a US contractor at $120 an hour for 1,800 hours is about $216,000; a dedicated MacroCoderz engineer is $2,800 x 12 = $33,600. The US hire is still the right call if that person will own the architecture and hire the next two engineers.
What does a first year actually cost? A worked example
Take the most common first AI project we see: search and answers over one body of company documents. Using only the published ranges above:
- Build: $5,500-9,000 for RAG over one corpus, 5-7 weeks.
- Running: $150-600 a month for embeddings, vector storage and inference, or $1,800-7,200 across twelve months.
- Year-one total: roughly $7,300-16,200, before any maintenance or new features.
The spread is almost entirely data quality and volume. If the documents are clean PDFs in one format, the project sits at the low end; scanned archives and several source systems push it to the high end. The larger, multi-tenant version of this system, built end to end, is written up in the Klebbix hybrid retrieval case study.
How can you reduce the cost of AI development?
- Start with the hardest part, on real data. Data quality and retrieval accuracy show themselves within days of touching the actual corpus.
- Cut integrations from version one. Each one you defer removes roughly a week and a failure mode.
- Let the first version suggest rather than act. Adding an approval step is far cheaper than building the safety net for full autonomy.
- Buy when an off-the-shelf product does 80% of what you need. Build when the value is in your own data or workflow - and do the two-year arithmetic, not the first-invoice one.
- Design for running cost during the build: a query cache and small-model routing are a few days of work.
- If you are not sure what to build, pay for a written readiness audit ($2,000-3,500) before committing to a build.
How do you read an AI development quote?
A lower quote is often a different scope, not a better price. Ask every vendor the same questions and compare the answers, not just the totals.
- Is evaluation included? A labelled test set and regression suite is 15-25% of a serious build. A quote without it is a demo.
- Which integrations are counted? Get the list in writing. If it grows from two to five during discovery and the timeline does not, something will give.
- What data condition does it assume? A vendor who has not seen ten representative files is guessing.
- Are model and infrastructure costs in the number? Ask whether API credit is resold with a markup or billed to your own accounts at cost.
- Is the price fixed, and what happens on overrun? On a true fixed price, underestimation is the vendor's problem; scope changes should be quoted in writing before work starts.
- Who owns the code, prompts and evaluation data at handover?
- What does maintenance cost after launch, and who does it?
Frequently asked questions
With MacroCoderz, fixed-scope AI builds start at $4,500 and most land between $6,000 and $26,000. A single-purpose chatbot is $4,500-6,000; document search over one corpus is $5,500-9,000; a custom agent platform is $14,000-26,000. US agencies typically bill $150-300 an hour for the same engineering, which is 3-5x the cost.
Most automation runs $60-250 a month, a single-corpus RAG system $150-600, and a production agent $200-3,000. The main components are model API spend and cloud hosting, which we bill to your own accounts at cost. Token volume and cache hit rate move the number far more than the model you pick.
For a defined outcome, a fixed-price project is cheaper. For an ongoing roadmap, compare fully-loaded annual costs: roughly $254,000 in year one for a US hire, $216,000 for a US contractor, and $33,600 for a dedicated MacroCoderz engineer at $2,800 a month. Hire in-house when the role will own the architecture for years.
Integration count. Every external system is an auth flow, a rate limit, a data model and a failure mode to handle. One integration is about a week; five is more than five weeks, because they interact. Data quality is second, and it is the most common reason an estimate doubles.
Usually because one includes evaluation and one does not, or because one assumed clean data. An evaluation harness is typically 15-25% of a build and a compliance review adds 20-40%. Ask each vendor what integrations, data condition and test coverage the number assumes.
Around $4,500. Below that, scoping and handover overhead outweighs the build, and a freelancer at $60-150 an hour is usually the better choice. A written AI readiness audit at $2,000-3,500 is the cheaper first step if you are not sure what to build yet.
Get a real number for your project
A free strategy session, a written scope and a fixed quote - based on your data, not a demo set.