How Much Does Custom AI Development Cost in 2026?

A model may cost a few cents per answer. A system allowed to read your CRM, verify an order, respect permissions, avoid inventing financial promises, and still work properly at 9 a.m. on Monday is no longer “a few API calls.”

Search for the cost of an AI agent in 2026 and you will find everything from $8,000 to $500,000 and beyond. That does not necessarily mean somebody is lying. “Agent” is being used for both a no-code workflow and a production product with five integrations, role-based access, audit logs, human approval, and an on-call support path.

This guide therefore gives you a budgeting model, not one magic number. The ranges below are editorial planning ranges for custom development aimed at European small and mid-sized businesses in 2026, excluding VAT. They are not a rate card or offer. A real estimate follows the scenarios, data, integrations, volume, and acceptance criteria.

The short answer: six budget levels

The lower boundary assumes a narrow scenario, existing systems, clean data, usable APIs, and quick client decisions. The upper boundary means several roles or channels, complicated exceptions, sensitive data, guaranteed latency, significant volume, or actions with consequences.

If employees only need individual assistance, test an enterprise AI tool or established SaaS first. Custom work makes sense when the value lies in your process, data, rules, or customer experience—and a ready-made product cannot reproduce it.

The first-year cost formula

Year 1 TCO = discovery + implementation + integrations
             + evaluation + security + rollout
             + 12 × (model usage + infrastructure + vendors
                     + human review + monitoring + maintenance)

Your budget needs two figures: cost to build and cost to own. A €20,000 project that costs €6,000 every month is more expensive in year one than a €45,000 project costing €1,000 monthly. A cheap demonstration may also require a near-total rebuild before it becomes production software.

Seven factors that move the estimate

1. The number of scenarios—not the model

“Where is my order?” and “Recommend a product, check stock, change my address, and issue a return” are not one chatbot. Every intent brings sources, rules, tests, exceptions, and escalation ownership. Three narrow scenarios are often cheaper than one vague “universal assistant.”

2. The condition of the data

If your return policy exists in three conflicting PDFs, product IDs are unstable, and exceptions live in managers’ heads, development begins with operational clean-up. AI does not reconcile the truth automatically. It scales inconsistency faster.

3. Integrations and API quality

A documented public API with a sandbox and webhooks may take days. A legacy ERP with custom authentication, incomplete data, and no test environment may take weeks. “Connect the CRM” is not an estimation unit. Each object, direction, permission, and failure path must be scoped.

4. Answering versus acting

A draft can be corrected before it causes harm. Sending an email, altering a booking, or issuing a refund requires authorization, idempotency, approval, limits, audit, and rollback. This is usually the largest jump in budget.

5. The required quality level

An internal summary can remain useful after edits. Medical, financial, or legal recommendations have different thresholds and risks—and may be unsuitable for autonomy. The more expensive the mistake, the more evaluation and control cost.

6. Volume, latency, and traffic peaks

One thousand documents processed overnight and one thousand simultaneous voice calls need different infrastructure. Real-time operation, long context, retrieval, audio, images, retries, and deeper reasoning all change variable cost.

7. Ownership and compliance requirements

Single sign-on, roles, processing region, audit, deletion, contractual controls, penetration testing, and incident procedures are real work. They do not look impressive in a demo, but they determine whether the system can be used.

Where the implementation budget goes

Model / prompts5–15%

Data / retrieval10–20%

Integrations20–35%

Product / UX10–25%

Evaluation / security15–25%

Launch / ops10–20%

Discovery maps the process, scenarios, sources, actions, risks, metrics, and technical unknowns. Data and evaluation cover source preparation, version control, real test cases, edge cases, and regression. Integrations and product cover APIs, authorization, state, user and reviewer interfaces, errors, and analytics. Production hardening adds logs, monitoring, cost limits, fallbacks, incident response, documentation, and handover.

If a proposal prices only “model configuration” and the chat frontend, the other work does not disappear. It arrives later as change requests or remains your operational risk.

Monthly costs after launch

Model providers publish prices per million tokens, tools charge per call, speech may charge per minute, and other services charge per user or task. Those rates change. Your estimate needs a formula, a volume assumption, and an alert when reality diverges—not one fixed-looking number.

A worked API-cost example

Imagine a support workflow handling 20,000 conversations each month. On average, every conversation sends 8,000 input tokens—the rules, history, and retrieved records—and produces 1,000 output tokens. For illustration, use €2.50 per million input and €15 per million output.

Input:  20,000 × 8,000 = 160M tokens × €2.50 / 1M = €400
Output: 20,000 × 1,000 =  20M tokens × €15.00 / 1M = €300
Base model cost: €700 / month

Add 30% for retries, evaluation calls, and uneven traffic:
Planning model cost: approximately €910 / month

This is an illustration, not a forecast or vendor quote. OpenAI’s current pricing page alone shows input and output rates varying several times between models, with batch and priority modes priced differently. Google and AWS likewise document tiers, caps, and processing options.

Now add human review. If 25% of conversations need three minutes from a support agent, that is 250 hours per month. At a fully loaded €20 per hour, review costs €5,000—more than five times the model bill. Optimising only tokens may be targeting the smaller line item.

Three worked project budgets

Ecommerce support copilot

Email, two languages, order status, delivery, and return-data collection; 2,500 contacts per month; a person sends the answer.

Internal B2B research assistant

A manager uploads approved material and receives a structured, cited briefing. The system does not write to CRM or contact a customer.

Invoice-processing agent with actions

Reads an invoice, verifies the supplier and purchase order, prepares a record, and routes exceptions. Creating the financial record requires approval.

These examples explain the mechanism; they do not promise another business the same figure. Replacing one integration or changing autonomy can move an estimate by tens of percent.

How to read a suspiciously cheap proposal

A low proposal can be excellent when it deliberately sells a narrow prototype. The problem begins when its language promises production. Check whether it includes:

• real acceptance cases, not only a happy-path demo;
• a separate estimate for every integration and permission;
• work on conflicting or outdated sources;
• identity and authorization checks outside the model;
• a human queue, stop conditions, and escalation ownership;
• logs, cost caps, monitoring, and failure notification;
• duplicate, timeout, unavailable API, and partial-action handling;
• regression tests after model or policy changes;
• production deployment, documentation, and handover;
• monthly third-party costs and support terms.

Ask what happens when the CRM is unavailable, the model returns an invalid format, or a user presses “execute” twice. “We will add that later” tells you the current product boundary.

Reduce cost without building something fragile

1. One workflow instead of a universal agent.
2. Read-only before write. Let the system retrieve and prepare before it acts.
3. Human review only where it earns its place.
4. Clean and assign data ownership before the contract.
5. Bring 50–100 real cases to discovery and evaluation.
6. Use an existing channel before building another interface.
7. Route between models and keep deterministic rules in code.
8. Use asynchronous or batch modes when latency is not valuable.
9. Write the out-of-scope list.

The most expensive way to economise is to skip evaluation and launch broad autonomy. A failure found in a test set costs hours. The same failure after hundreds of customers costs data repair, support, reputation, and sometimes legal work.

Copy-ready budget brief

BUSINESS OUTCOME
Move ______ from ______ to ______ without worsening ______

FIRST WORKFLOW
User / input → result: ______
Monthly volume / peak / required latency: ______

DATA AND INTEGRATIONS
Sources and owners: ______
Read systems / write systems: ______
Actions requiring approval: ______

QUALITY AND RISK
50–100 acceptance cases ready: yes / no
Zero-tolerance failures / escalation: ______
Logs / retention / roles: ______

PRICE SEPARATELY
Discovery; build; each integration; evaluation; security;
production rollout; third-party cost at 1×, 2×, and 10×;
human review; maintenance; Year 1 TCO.

ASSUMPTIONS
Client provides: ______
Excluded: ______
Variable with the largest price impact: ______

Before requesting prices, complete the copy-ready AI assistant brief. It makes different vendors estimate the same product rather than three private interpretations.

When not to build

Buy when the process is standard, the data already lives in a popular platform, and 70–80% of a SaaS product is enough. Configure when interfaces and connectors exist but your workflow rules create the value. Build when you need proprietary data and permissions, unusual logic, several systems, controlled autonomy, or an AI feature customers pay for.

Wait when there is no process owner, baseline metric, data access, or agreement about correct work. Start with making the business AI-ready.

Frequently asked questions

How much does a simple AI chatbot cost?

Configuring an existing product may cost €500–€5,000. A custom proof of concept may cost roughly €4,000–€12,000. A production assistant with live data, integrations, evaluation, and escalation starts substantially higher.

How much does a business AI agent cost?

A narrow production agent that takes actions often warrants a €40,000–€120,000+ planning range, but systems, permissions, risk, volume, and support can push it far beyond that.

Why is the API cheap while development is expensive?

The product must gather correct context, authenticate the user, connect systems, constrain actions, handle failures, present the outcome, record it, and remain maintainable.

What budget is needed for an AI pilot?

Testing one narrow hypothesis may fit €4,000–€12,000. A live production pilot with integrations, security, and real users may require €15,000–€40,000 or more.

How much should we budget for maintenance?

Estimate model, hosting, monitoring, human review, and recurring support at 1×, 2×, and 10× volume. A universal maintenance percentage is less useful than a workload model.

Sources and notes

• OpenAI API pricing — current input/output, batch, priority, and tool pricing.
• Google Gemini API billing — usage tiers, spend caps, and billing components.
• Amazon Bedrock pricing — pricing by model and processing tier.
• The Crunch and MarsDevs — examples of how widely the market prices different products called “AI agents.”

The most useful number in a proposal is not the total. It is the assumption beside it: which workflow, data, permissions, volume, quality level, and failure behaviour. Once those are visible, price becomes a decision. Without them, it remains an impression.