How to Estimate Cost & Time to Build a Custom AI in 2026

If you’ve asked three different AI agencies for a quote, you’ve probably got three wildly different numbers — €15K, €60K, €240K. The reason isn’t shady salespeople. It’s that “custom AI” covers a 50× cost range, and most agencies don’t separate the cheap parts from the expensive ones.

This is the framework I use in every scoping chat: five inputs and a simple formula give a ballpark. It’s the same framework that powers my interactive cost estimator, so if you want the answer without doing the maths, start there.

Why most custom AI estimates are wrong

Three traps that produce bad numbers:

1. They price the AI, not the plumbing. The actual LLM-and-prompt work is typically 20–30% of the cost. The other 70% is integrations, admin panels, error handling, audit trails, deployment. An estimate that ignores the plumbing is an estimate of about a third of the project.
2. They use a unit price for “an integration.” An integration with Slack’s well-documented API is a day of work. An integration with a 1990s ERP that returns CSVs over SFTP is two weeks. Treating them as the same number is how projects 5× over budget.
3. They forget the data. If the AI needs to read from your knowledge base and your knowledge base is 4,000 PDFs in a SharePoint folder, you’re paying for ingestion, OCR and chunking before the AI does anything useful.

The framework below catches all three.

Step 1: Define the AI’s job in one sentence

Before you estimate anything, write the job down in plain English. Not the marketing version — the operational one.

Bad: “Use AI to be more efficient with customer support.”

Good: “Reply to incoming WhatsApp messages from retail customers in Arabic or English, look up their order in Shopify, and either answer the question or escalate to a human agent if confidence is low.”

The good version tells you:

• The channel (WhatsApp Business)
• The customer language (Arabic + English)
• One integration (Shopify) and a likely second (escalation route into your support tool)
• The success boundary (escalate when uncertain)

You can’t estimate the bad version. You can estimate the good version in five minutes.

Action: Write your one-sentence job description. If you can’t, you’re not ready to scope. You’re still in problem-discovery, and that’s where the scoping chat should start.

Step 2: Count the integrations — and grade them

List every external system the AI reads from or writes to. Then grade each one A, B or C:

Sum the days. That’s your integration effort. To turn it into money, multiply by a market rate: SFAI Labs publishes blended agency rates of $150–$350 an hour.

This is also where you discover dealbreakers. If an absolutely-required system grades C and there’s no plan B, the project shouldn’t start until the integration path is real. Better to find out now than three weeks in.

Step 3: Score the data — the most underrated cost driver

If your AI needs to know things, you need to grade the data it needs to know.

• Green data: queryable, structured, accessible (e.g. a database, a clean API). The AI just reads it. Cost: minimal.
• Yellow data: text and PDFs that exist but need ingestion (chunking, embedding, indexing). Cost: 1–3 weeks of engineering depending on volume.
• Red data: scanned documents, handwriting, photos, video, undocumented spreadsheets. Cost: OCR pipelines, manual labelling, extraction work. Frequently 30–50% of total project cost in regulated industries.

Most cost surprises in custom AI come from underestimating Yellow and missing Red entirely. That is why I walk through your data with you on the scoping chat, with the screen shared.

Step 4: Pick the deployment shape

Custom AI projects fall into four shapes. The shape determines roughly 60% of the cost. Pick yours. The range on each card is what a named company publishes for that kind of build, so you have an independent anchor (every source is linked in the 2026 pricing benchmarks):

If your project description fits two shapes, pick the bigger one for the estimate. Custom AI projects don’t shrink under pressure; they expand.

Step 5: Apply the formula

Combine the four inputs from Steps 1–4 with two multipliers:

base_range = (published market range for your shape, step 4)

integration_mult =
  1.0   if 1–3 systems, mostly grade A
  1.25  if 4–8 systems, or some grade B
  1.6+  if 9+ systems, or any grade C

hosting_mult =
  1.0   managed cloud (hosted for you)
  1.15  your server / on-premise
  1.3   compliance-heavy on-prem (HIPAA, SOC 2, MAS, APRA, etc.)

estimate_low  = base_low  × integration_mult × hosting_mult
estimate_high = base_high × integration_mult × hosting_mult

Add 10–30% for Red data work if applicable. Add 1–3 weeks of timeline if your data needs ingestion before the AI can do anything.

That’s your ballpark against the market. It is not a quote. I don’t publish a price list: every project I take on is a fixed price, quoted after a short scoping chat with me and a one-page technical spec, so you know the full number before anything starts.

Three worked examples

Each one runs a sample brief through the five steps. They illustrate the method; they are not client projects.

Example 1: Retail WhatsApp support bot

• Job: Reply to WhatsApp messages, look up orders in Shopify, escalate when uncertain.
• Integrations: WhatsApp Business API (grade A, 1d), Shopify (A, 1d), human escalation via Slack (A, 0.5d). Total: ~3 days.
• Data: Green, everything queryable.
• Shape: Chatbot.
• Multipliers: Integration 1.0, hosting 1.0 (managed cloud).
• Estimate: 4 weeks.

Example 2: Bank reconciliation + invoice automation

• Job: Pull bank statements daily, match incoming payments to outstanding invoices in AFAS, flag discrepancies for finance team.
• Integrations: Bank API (B, 4d), AFAS (B, 6d), Stripe (A, 1d), Slack (A, 0.5d). Total: ~12 days.
• Data: Green for live data, Yellow for historical statements (PDF).
• Shape: Automation.
• Multipliers: Integration 1.25, hosting 1.0.
• Estimate: 8 weeks, plus about 1 week of ingestion for the historical PDFs (Yellow data).

Example 3: Internal AI assistant for legal team

• Job: Search across 8,000 contracts and case files, answer questions with citations, draft first-pass redlines.
• Integrations: SharePoint (B, 5d), DMS (C, 12d), Outlook for delivery (A, 1d). Total: ~18 days.
• Data: Mostly Yellow (PDFs, Word docs), some Red (scanned legacy contracts).
• Shape: Custom platform.
• Multipliers: Integration 1.6, hosting 1.15 (on-premise legal hold).
• Estimate: 10 to 14 weeks, plus 20% for Red data.

The five mistakes I see weekly

1. Estimating before defining the job. Vague problems get vague (and wrong) numbers. Step 1 is non-negotiable.
2. Treating all integrations as equal. Grade them A/B/C. Pretend a C is an A and you’ll be wrong by 5×.
3. Forgetting the data. If you have 10,000 PDFs that need to inform the AI, that’s a project on its own. Cost it separately.
4. Picking the wrong shape. “A chatbot” that needs admin panels, multi-user access and audit logs is a custom platform. Don’t price it as a chatbot.
5. Skipping the scoping chat. A structured conversation narrows the estimate more than any formula can. There’s no shortcut that beats it.

If you’d rather not do the maths

That’s what the interactive cost estimator is for. Three questions, a first read in seconds, built on the formula above. If you want the written-down version with timelines, the timeline page covers the four phases week by week.

If you want a real number for your specific situation, that’s the scoping chat with me. It ends in a fixed price, not a range. By the end you’ll know the likely cost, the rough timeline and whether custom AI is the right tool, even if the answer is no.

Related reading

• Custom AI Development in 2026: Real Cost, Timeline & How to Pick a Developer — the comprehensive guide
• How Much Does Custom AI Development Cost? 2026 Pricing Guide — pricing-only deep dive
• How custom AI development actually works — six-stage methodology
• Questions to ask before hiring an AI developer