What AI Implementation Actually Costs in 2026

From the few of these we have deployed, the spread between a cheap and an expensive build is rarely about the AI. It is about how many other systems the AI has to touch. A chatbot that answers questions from a single PDF using Claude Sonnet 4.5 over the Anthropic API can ship in two weeks for under $30K. The same chatbot that has to read your Zendesk tickets, write back to HubSpot, escalate to a human in Slack and stay compliant with GDPR is six months of work and starts at $250K.

Both numbers can be defensible. Neither tells you which one fits your business. The cost of the LLM call is usually under 15% of either total. Everything else is plumbing, evaluation, security review and the awkward weeks where the product team and the data team argue about who owns the prompt.

The ranges below come from the work we have shipped plus rough numbers I have heard from operators in the past year. They assume a partner who has built this category before. If you are doing it fully in-house with no prior LLM production experience, the dollar cost drops and the calendar cost roughly doubles. The pattern I keep seeing is that the cheap quotes leave out integration and the expensive quotes leave out the ongoing API spend. The honest version itemizes both. If you are also weighing which type of system to build, the breakdown in agents vs chatbots vs copilots is worth a read first.

MIT's Project NANDA reported in July 2025 that the GenAI deployments which did pay off cited 2 to 10 million dollars a year saved on customer service and document processing, the unglamorous end of the stack.

At the moment, Claude Sonnet 4.5 sits at $3 per million input tokens and $15 per million output on the Anthropic pricing page. GPT-5 and GPT-4o land in a similar zone per the OpenAI pricing page. Those numbers sound trivial until you do the multiplication. A support agent that handles 5,000 conversations a month with an average of 8,000 input and 1,500 output tokens per conversation runs about $120 input and $112 output, so $232 a month. The same agent at 200,000 conversations is closer to $9,300 a month. Same prompt, same model, three orders of magnitude difference in invoice.

Vendor quotes that bury this number inside a flat monthly fee are usually pricing for low volume and will renegotiate the moment you scale. Talking to operators in production, the same pattern repeats. Month three is fine, month nine the invoice has tripled and nobody warned them. If your vendor quote does not itemize token spend separately from build cost, walk away. That is where projects bleed for years.

Beyond the LLM API, every production AI system carries a small zoo of supporting bills. A typical mid-market support agent runs something like Pinecone serverless for vector storage at $50 to $300 a month, a Hetzner CCX33 dedicated vCPU box for the orchestration service at roughly €60 a month, Vercel Pro for the frontend at $20 per seat, an observability tool like Helicone or LangSmith at $50 to $500 a month depending on volume and a managed Postgres for session and audit logs at $25 to $200. None of these are big on their own. Together they easily clear $1,000 a month before the LLM even shows up on the invoice.

If you are running heavy inference on open-weight models, Modal, Replicate and Fireworks all expose hosted GPU pricing that is worth a look before committing to a self-managed cluster. For a useful outside read on why aggregate infrastructure spend keeps creeping up, the a16z economic case for generative AI is a fair starting point.

Operators love to ask whether they should switch from Claude to GPT-5 to save money. In practice, the difference between picking the right model and the wrong model on a per-call basis is maybe 30 to 50%. The difference between a sloppy prompt that ships 12,000 tokens of redundant context every turn and a tight one that ships 2,000 is a clean 6x. The fastest way to cut your monthly invoice is almost always to rewrite the prompt, cache the system message and trim the retrieval context, not to migrate providers. I have watched teams spend six weeks on a model migration that saved 20%, when an afternoon of prompt cleanup would have saved 70%.

So pick a provider that supports prompt caching (Anthropic and OpenAI both do now) and structure your system prompt so the cached prefix is huge. Pricing pages list cache reads at a fraction of the input cost. That is the single biggest knob on the run-cost side. The OpenAI cookbook has a useful set of patterns for thinking about this.

If you want to sanity-check whether that math pays back in your specific business, the interactive ROI calculator takes your conversation volume, agent salary and resolution rate as inputs and stress-tests the case.

Deloitte's global automation survey, published in June 2022, found organizations that scaled automation cut costs by 32 percent on average, while the typical program took about 22 months to pay back.

If somebody asked me to spend their money tomorrow, the order of operations would be to ship the smallest version of the use case in four weeks against an off-the-shelf API, instrument it heavily from day one so you actually see token cost per conversation and per intent, then make the build-vs-buy call at month two with real data. Almost every team that spent $400K on a custom build before doing this step ended up rewriting most of it within a year.

The second move is to put a six-month review on the calendar before the contract is signed. Run cost will not look like the proposal said it would. Either volume undershot and you are paying for capacity you do not need, or it overshot and the per-call price needs renegotiation. Both are fine outcomes if you planned to revisit them. Neither is fine if you only notice when the CFO asks why the AWS bill doubled.