How much does AI implementation cost for a mid-market business

A mid-market business with 50 to 500 staff should plan on roughly US$1,100 to US$2,000 per employee in year one of AI implementation in our labelled model: US$100,200 at 50 staff, US$228,600 at 150 and US$558,000 at 500. Seat licences, priced from Claude’s published US$20 a seat a month, are only 7% to 19% of that total.

• The question, as people ask it: “how much should I budget for AI across my whole company, not just one project?”
• The named framework: the three-layer budget. Seats (licences for staff), workflows (custom agents that do a job end to end) and guardrails (training, an owner, policy and security review).
• The biggest layer: workflows, 39% to 55% of year one. The people lines in the guardrail layer, training and an owner, are 33% to 38%.
• The decision rule: the 92-hour rule. A US$55,000 workflow must save about 92 staff hours a month at US$50 an hour, or earn the equivalent, to pay back in year one.
• How this differs from a single-project quote: our page on what one AI implementation costs across its five invoices prices one workflow in depth. This page sizes the whole company’s budget by headcount.

How much does AI implementation cost for a 50, 150 or 500-person business?

AI implementation for a mid-market business costs about US$100,200 at 50 staff, US$228,600 at 150 and US$558,000 at 500 in year one, in our assumption model.

Every quantity below is an assumption you should replace: 60% of staff get an AI seat; workflows grow from one to four with headcount; each workflow costs US$25,000 to build plus US$2,500 a month to run and maintain; training is six hours per seated person at a loaded US$50 an hour; an internal AI owner is 0.2, 0.5 or 1.0 of a US$120,000 loaded role. Only the seat price is published, on Anthropic’s Claude pricing page, read 15 October 2026.

The per-employee figure falls with size because workflows and the owner role are shared across more people, while seats scale one for one. In a mid-market AI budget, the licence line is the one everybody quotes and the one that matters least.

The three-layer budget: seats, workflows and guardrails

The three-layer budget splits mid-market AI spend into seats, workflows and guardrails, because each layer is priced in a different unit and owned by a different person.

• Seats are priced per person per month and owned by IT. They make individuals faster at writing, analysis and research, and the saving is spread thinly across many people.
• Workflows are priced per build plus a monthly run cost and owned by the head of the function. A workflow does a whole job, for example replying to every new lead and booking a call, and its return can be measured directly.
• Guardrails are priced in people’s hours and owned by whoever is accountable for AI risk. They cover training, a named owner, an acceptable-use policy, and a security and privacy review before customer data touches a model.

Budgets that fund only the first layer buy activity without a measurable result; budgets that skip the third find out after launch. For the post-launch costs that sit inside the guardrail layer, such as staff time spent checking AI output, see the hidden costs of AI projects.

What do AI seat licences cost a mid-market company?

AI seat licences for a mid-market company cost US$20 a seat a month on Claude’s Team plan billed annually; Enterprise adds usage at API rates.

The US$10 a seat a month usage figure in the 500-staff column above is our assumption, not a published price: on usage-based billing it depends entirely on how much your staff run the models. Other vendors also sell per-seat plans; price yours from each vendor’s own page on the day you budget, because these prices change.

The 92-hour rule: when does a custom AI workflow earn its place?

The 92-hour rule says a US$55,000 first-year AI workflow pays back only if it saves about 92 staff hours a month at a loaded US$50 an hour.

The formula: (build + 12 months of run cost) ÷ 12 ÷ loaded hourly cost of the people whose time it replaces. US$55,000 ÷ 12 is US$4,583.33 a month; at US$50 an hour that is 91.7 hours. For a revenue workflow, swap hours for gross profit: the workflow must add US$4,583.33 of gross profit a month.

Rank candidate workflows by this number before choosing which to build. If no candidate clears it, the first year’s budget belongs in seats and guardrails, and that is a legitimate answer.

Why are sales and marketing workflows often funded first?

Sales and marketing workflows are often funded first in mid-market AI budgets because their output, booked calls and closed revenue, is already counted in the CRM.

A support or finance workflow pays back in hours saved, which someone has to estimate. A lead-response or appointment-setting workflow pays back in booked calls, which the calendar records. That makes it the easiest line in the three-layer budget to defend at a year-one review. LeadsNow sees booked-call costs of $30–$400, depending on industry, offer, price and many other variables, so divide any sales workflow’s monthly cost by the calls it books and compare.

A sales workflow is also the one layer that can move from fixed to variable cost. Instead of a US$25,000 build and a monthly run cost, a done-for-you service such as LeadsNow’s AI appointment setting charges 5–25% of the revenue we generate for you (revenue share), or an equivalent pay-per-appointment fee, or a mix of both, set by lead volume, what you sell and its price, the type of business, and how much of the funnel we run. That suits companies without an engineer to own a workflow; it is the wrong choice where the workflow itself is intellectual property you plan to sell. Our B2B lead generation guide covers the wider pipeline the workflow feeds.

What does a mid-market AI budget leave out?

A mid-market AI budget built this way leaves out data clean-up, integration surprises and productivity dips, each of which belongs to a specific workflow rather than the company total.

• Data remediation: priced per source system once a workflow is chosen; it can be large if your CRM is messy.
• Integration rework: CRM upgrades and renamed fields re-open finished work.
• The first-months dip: people slow down while they learn a new process.
• Marketing tools with AI built in: often already in the marketing budget; see what AI marketing costs as a share of revenue.

Add these per workflow, not as a company-wide percentage, because they vary more between two workflows than between two companies.

Frequently asked questions

How much does AI implementation cost for a mid-market business?

In our labelled model, about US$100,200 in year one at 50 staff, US$228,600 at 150 and US$558,000 at 500, or US$1,116 to US$2,004 per employee. Seat licences are 7% to 19% of it; custom workflows and the people who run them are most of the rest.

What does an AI seat cost per employee?

Anthropic’s Claude pricing page, read 15 October 2026, lists the Team plan for teams of 2 to 150 at US$20 per standard seat a month billed annually (US$25 monthly), and Enterprise at US$20 per seat a month plus usage billed at API rates.

How many AI workflows should a 150-person company build in year one?

Only as many as clear the 92-hour rule. Our model assumes two at 150 staff, but if no candidate saves about 92 staff hours a month at US$50 an hour, or adds the equivalent gross profit, spend the year on seats, training and an owner.

Should AI implementation cost be budgeted per employee?

Seats and training should be; workflows should not. Seats scale with headcount, but a workflow costs the same whether 50 or 500 people benefit, which is why per-employee cost falls from US$2,004 to US$1,116 across our three sizes.

Who should own the AI budget in a mid-market company?

Split it by layer: IT owns seats, each function head owns their workflows and their payback, and a named AI owner holds the guardrails. Our model costs that owner at 0.2 of a role at 50 staff rising to a full role at 500.