What drives the cost of an AI implementation, and how to estimate it

The first question about AI at work is usually what it will cost. The subscription is easy to look up, and it is rarely the number that decides the budget. What decides it is visible before anyone sends a quote: the workflow, the systems around it and the people who keep it running.

The short answer

An AI implementation has three cost lines: the tools, the work of building and connecting the workflow, and your own team's time. The tools you can price yourself from vendor pages. The build and the time depend on how many systems the workflow touches, how clean the data is, how many exceptions and approval steps it needs, how much volume runs through it, and who maintains it. One workflow that drafts and waits for a person's approval, on tools you already pay for, sits at the low end. Whoever quotes the work, a consultant or an implementation partner, should break the quote down by these same drivers.

What drives the cost

• Tools: seats and usage. Seat plans are predictable. Usage plans, billed in tasks, credits or tokens, grow with volume and with settings.
• Systems the workflow touches. Each connection, such as email, a CRM, accounting software or a shared drive, adds setup, permissions and testing. One system is a small job. Four is a project.
• Data quality. If the inputs are scattered across inboxes, inconsistent or stuck in scanned PDFs, cleanup comes before anything else.
• Exceptions and approvals. The common case is quick to build. The edge cases take the time, and every step that needs a person's approval has to be designed and tested.
• Language and format. Hebrew documents and mixed Hebrew and English messages need real examples in testing from the first week.
• Upkeep. Vendors change features, prompts drift and people leave. Someone has to own the workflow after launch.

A worked example

Example, for illustration. A 10-person insurance agency wants incoming quote requests turned into draft replies that an agent approves, with an AI step inside a Zapier automation, as in our note on Zapier's AI step.

Usage first. By Zapier's formula, an AI step on the Advanced tier (3 tasks per unit) with one price list lookup uses 3 + 3 = 6 tasks, and the draft and a log row add 1 task each, so about 8 tasks per request. At 100 requests a month that is 800 tasks, just past the 750 task tier, so the agency needs the 1,500 task Professional tier at $39 a month billed annually, per Zapier's pricing page. At 200 requests it is 1,600 tasks and the 2,000 task tier at $49. Left on the Premium default (5 tasks per unit), the same 200 requests use 12 tasks each, or 2,400 tasks, past that tier.

Seats next. If three agents also draft with Claude on Team Standard seats, Anthropic's pricing puts that at 3 × $25 = $75 a month on monthly billing, or 3 × $20 = $60 billed annually, before tax.

Then time. Assume 16 hours of the agency's own time in the first month, for mapping the steps, writing the rules and testing on 30 past requests, and 2 hours a month after that. At an assumed internal cost of 150 NIS an hour, that is 2,400 NIS of time in the first month and 300 NIS a month after. Write this line down: it is the easiest one to leave out of a budget.

The build itself, connecting the inbox, the price list and the approval step, scales with the drivers above, which is why this example gives it no number.

How to estimate it, step by step

1. Pick one workflow and count it: requests per month, steps, and the people who touch it.
2. List the systems it touches and who controls access to each one.
3. Price the tools from the vendor pages: seats times the seat price, and usage as volume times units per run. Check what happens past the limit.
4. Pull 20 to 30 real examples, including messy and Hebrew ones, and count the exceptions.
5. Estimate your team's hours for mapping, rules, testing and the first month, plus monthly upkeep, and put a stated internal cost on them.
6. Compare quotes for outside help line by line against this list: the workflow, the systems, the data work, the approvals and who maintains it.

Mistakes that raise the cost

• Starting with several workflows at once, so nothing is finished and every cost is shared.
• Pricing the seat and ignoring usage. Defaults matter: new AI by Zapier steps on paid plans start on the Premium tier.
• Skipping the data check and finding out mid-build that the inputs live in scanned PDFs.
• Automating the old process as it is. In McKinsey's survey from March 2025, the redesign of workflows had the biggest effect on bottom-line impact from gen AI, out of 25 attributes tested, yet only 21 percent of respondents said their organizations had fundamentally redesigned at least some workflows.
• Leaving review undefined. The same survey found that 27 percent review all gen AI content before it is used. Decide in advance what a person checks.

When outside help is worth it

One workflow, one or two systems, clean inputs and a clear owner: most teams can do this themselves with the steps above. Outside help pays off when the workflow touches several systems, when the data needs cleanup, when approvals and permissions need design, or when nobody has the hours. The AI implementation page describes how that work runs, one process at a time. If it is not yet clear which workflow comes first, the AI consulting page covers the mapping that comes before it.