What AI implementation really costs — the four drivers behind the number

AI implementation is five invoices, not one: licence, build, integration, data remediation and run cost. Gartner’s 29 July 2024 release put deployment approaches used to transform a business model at US$5m–US$20m; separately, at its October 2024 Symposium Gartner said organisations that deployed AI spent US$300,000–US$2.9m on the proof-of-concept phase alone in 2023. The model licence is almost always the smallest line on the page.

• Published ranges: US$300,000–US$2.9m on the proof-of-concept phase alone, by organisations that deployed AI (Gartner, 2023 spend, reported October 2024); US$5m–US$20m for business-model-scale deployment (Gartner, 29 July 2024).
• Four drivers move your number: systems touched, state of your data, how many people’s jobs change, and monthly volume after go-live.
• The line everyone forgets: integration — CRM write-back, telephony, calendar, identity, reporting, staging, security review — priced per system and per environment.
• The blow-out: Gartner told its 2024 Symposium audience CIOs could miscalculate AI costs by as much as 1,000% as they scale.
• The commitment: RAND advises committing a product team to one problem for at least a year before starting.

How much does AI implementation cost? The published ranges, and what each one measures

The figures in circulation differ by two orders of magnitude because they measure different things. Side by side with what each excludes, they stop contradicting each other. Any AI cost benchmark you quote to a CFO needs its scope stapled to it, or the first analyst who reads it will find the gap.

Two are forecasts, three are surveys, and none is a quote for your business — which is why the rest of this page prices the drivers instead.

The four drivers behind the number

Scope, data, people and volume. Every quote that surprises you later is one of those four left off. You cannot price an AI implementation from the use case alone — you price it from the systems the use case has to touch.

The integration cost everyone forgets

The rule we apply to any AI budget we are shown. The five-invoice rule: a credible AI implementation budget has five cost lines on it — licence, build, integration, data remediation and run. If your business case has fewer than five, you are holding a licence quote, not a budget.

Integration goes missing because it is invisible until someone tries to write a record back. On one AI qualification and booking workflow it means: CRM field mapping and write-back with an audit trail; telephony or messaging provisioning with consent and opt-out flags; calendar access and routing rules per rep; single sign-on and role permissions; a staging environment that mirrors production well enough to test against; logging, retention and reporting; and a security review before any of it touches live customer data. For regulated outbound, add consent capture and suppression — every item in our AI outbound compliance checklist for enterprise teams is integration work, and all of it is billable.

It is also the only line charged twice: built once, then re-opened whenever your CRM upgrades, a telephony provider changes an API, or operations renames a field without knowing the agent reads it.

What the data work costs when your CRM is the input

An AI agent reading your CRM inherits every decision your team has made about it. Price the data work by counting five things before anyone quotes: source systems holding contact records; duplicate share; share with no valid phone or email; share with no consent status; and custom fields the agent must read to decide. Those five counts are the data quote, and no vendor can produce them for you because they cannot see inside your instance. Our CRM data hygiene framework sets out the audit behind them. Data remediation is the only AI cost line you can cut to near zero with your own labour, which is why it is the first to scope and the last anyone does.

Change management is a cost line, not a slide

Deloitte’s State of AI in the Enterprise 2026 (3,235 leaders, 24 countries, surveyed August–September 2025) found 40% of surveyed organisations reporting cost reductions and 20% reporting increased revenue. A separate question in the same survey splits organisations by how far AI has changed the way they work: 34% are starting to use AI to deeply transform, 30% are redesigning key processes around it, and the remaining 37% are using AI at a surface level, with little or no change to existing processes. The two questions have different denominators and do not net against each other — but that last third is the population that bought the licence and never funded the change management line.

Price it in manager hours, not licences: rewriting the process the agent now performs, retraining the people whose queue it feeds, QA listening in the first eight weeks, and amending the incentive scheme when an agent books meetings a rep did not source. Stanford HAI’s 2026 AI Index reports 88% of respondents in 2025 saying their organisation uses AI in at least one business function, up from 78% in 2024 — a self-reported figure the Index republishes from McKinsey’s 2025 State of AI survey, not a count of firms. Adoption is no longer the differentiator, so the return goes to organisations that changed the process rather than the ones that bought the tool.

What it costs every month after go-live

Run cost scales with volume, and volume grows after launch. It is inference or per-minute charges, monitoring, human escalation for conversations the agent should not finish, re-testing when a model version changes, and transcript storage and retention. Our breakdown of what AI voice agents cost per minute in the US shows how a headline per-minute rate becomes a loaded minute once failed dials are counted; the same arithmetic applies to any per-unit AI price you are quoted.

Budget run cost at your expected volume twelve months after launch, not at pilot volume. That single substitution is most of the gap Gartner described when it told its 2024 Symposium audience that CIOs could miscalculate AI costs by as much as 1,000% as they scale.

Worked example: costing one AI appointment-setting workflow end to end

Every figure below is a placeholder. The arithmetic is the asset, not the numbers.

1. One-off block. Build 60,000 + integration 70,000 (four systems, two environments) + data 30,000 + change management 20,000 = 180,000.
2. Amortise over 12 months. 180,000 ÷ 12 = 15,000 per month.
3. Add run cost at year-one volume, not pilot volume: 9,000 per month.
4. Monthly total: 15,000 + 9,000 = 24,000.
5. Divide by booked, qualified meetings, not conversations. At 120 a month: 24,000 ÷ 120 = 200 per booked meeting in year one, falling to 75 in year two once the one-off block is paid off.
6. Compare against the same denominator you use today, then multiply by show rate and close rate to get cost per closed deal — the number the CFO is actually approving.

The trap is the denominator: a cost per conversation looks excellent and means nothing, because conversations do not appear in revenue. Show rate is the step most models omit; it varies by offer and reminder cadence, and reaches up to 93% on our best-performing accounts.

When to build it, and when to buy the outcome instead

The crossover is volume and engineering capacity, not enthusiasm. These thresholds are our own decision rules, not a survey finding. Read the rows together rather than the volume row alone: with no engineering capacity to hold for twelve months, buying the outcome is the right call at any volume.

Buying the outcome is how our own enterprise lead generation service is priced: a revenue share of 5–20% of the sales we help generate, or roughly 1–5% of closed-deal value per appointment, paid on booked qualified appointments rather than retainers or seats. That moves the four drivers off your balance sheet, and it is the wrong answer when the workflow is intellectual property you intend to sell. Our published methodology defines the 7x average sales lift as trailing three-month closed-deal revenue at month six against the three months before launch, and discloses on the same page that the median is closer to 4x — use the median in the business case.

Frequently asked questions

What will AI implementation cost for my business in the first year?

Price the five invoices separately, then amortise the one-off block over twelve months and add run cost at year-one volume. For scale reference, Gartner’s 29 July 2024 release, reprinted in full, put deployment approaches used to transform a business model at US$5m–US$20m, and Gartner analyst Rita Sallam said in the same release that “there is no one size fits all with GenAI, and costs aren’t as predictable as other technologies”. A single workflow sits far below that range.

How much of an AI budget is the AI itself?

Usually the smallest share. Licence and inference are metered and quotable; integration, data remediation and change management are bespoke to your systems and your people, and all three are priced per system or per head rather than per token. If a quote contains only the licence line, the remaining four invoices have not been removed — they have been deferred to you.

How long before an AI implementation pays for itself?

Longer than a pilot suggests, because the one-off block is spent before the first result. RAND’s 2024 study of AI project failure, based on interviews with 65 experienced data scientists and engineers, recommends that leaders be prepared to commit a product team to a specific problem for at least a year, and warns that projects not worth that commitment are probably not worth starting.

What is the most common reason an AI budget blows out?

Scaling assumptions. Gartner told its 2024 Symposium audience, as reported by CIO Dive, that CIOs could miscalculate AI costs by as much as 1,000% as they scale, and that organisations deploying AI spent between US$300,000 and US$2.9m in the proof-of-concept phase alone in 2023. Pilot-volume run cost multiplied by production volume is where the error compounds.

Can we avoid the implementation cost entirely?

Only by changing what you are buying. If you buy an outcome — booked qualified appointments, on a revenue share or per-appointment basis — the build, integration and run costs sit with the provider and you pay against results. You still own the data work, because nobody can clean your CRM from outside it.