Planning tool · Not a proposal

What should your AI programme cost?

A planning tool for your problem definition and solution exploration phase. Use it to frame the investment, align your buying group, and walk into vendor conversations with the right questions, before anyone has scoped anything.

Kowalah

AI Programme Cost Estimate

Indicative planning estimate · Generated 29 August 2026
Configuration: 1,000 employees
Start here

Sketch the programme you’re considering

Every field is an assumption, not a commitment, and each figure below updates as you change it. Nothing is sent to us or to anyone else, it stays in your browser and in the link you can copy at the foot of the page.

Your Organisation
1,000employees
2005001k2.5k5k10k20k
$200ma year
$13.7m$100m$1bn$6.8bn

Used to size the value at stake in each domain you pick, and to show what the programme is worth as a movement in EBITDA. That is $200k per employee at your headcount, which is a normal range for an organisation this size.

The currency toggle sets every figure on this page, converted at static rates. AI platform licensing is billed in USD; professional services are quoted in your delivery partner's local currency.

Which processes are you transforming?

Organisations run 1020 process domains. We’d recommend a first programme targets the one or two with the highest value and the best chance of landing. Focused programmes deliver, broad ones stall.

Operating processesWhere the commercial prize sits: revenue, customers, delivery.
Management and supportEfficiency and cost plays. Valuable, but rarely the biggest prize.
How that becomes a number of agents
1domain×2solutions each×4agents each

A solution is a system inside the domain, such as customer service automation, demand forecasting or campaign production. The agents are the individual capabilities that make it work, and a solution needs 3 to 5 of them to move one business metric. A full domain transformation lands at 5 to 15 agents in total. If you don’t know yet, leave this alone; it’s the sort of thing a discovery workshop settles in an afternoon.

8agents in total
6live by the end of year one

Year one doesn’t finish this scope, and that’s normal. The rest follows as the programme continues. If you need more live sooner, the lever is fewer domains rather than a faster build.

What you’d be trying to move
Market & sellRevenue per repCost of customer acquisitionWin rateSales cycle length
Regulated industry?
Finance, healthcare, legal, or similar. Adds governance and acceptable-use work to Strategy, and a security review and evidence trail to every build, so it moves two lines, by different amounts.
Someone in-house to own this?
An AI owner with the remit to run the programme and, more importantly, to keep the agents running once your delivery partner has gone. Without one you stay dependent for longer, which shows up in the ongoing line rather than the build.
The bottom line

What your programme costs

A complete AI programme has six budget lines, and they are not all the same kind of money. Three are one-off, two recur every year, and one never leaves your own payroll.

Year one, one-off$1.2m–$1.6mProfessional services from your delivery partner. Paid once.
Pillar 1
Strategy & Advisory
$78k–$164k
The work before a line of code: mapping the domain, agreeing the roadmap and the business case, and setting up governance you can report to a board. Runs once per domain, so it does not multiply with the number of agents.
Pillar 2
Design & Build
$1.1m–$1.4m
Designing, building, testing and deploying the agents themselves, 8 of them at $134k–$173k each. The largest single line in the budget, and the one that scales with how much you choose to build rather than with how many people you employ.
Pillar 3
Change Enablement
$19k–$23k
Champions, leadership coaching, training and adoption measurement, running alongside the build from week one, not bolted on at the end. Scales with the people the agents reach, not your whole headcount.
Ongoing, every yearThe run rate after year one. Two different payees.
Pillar 4
Managed Services
$84k–$253k/yr
Ongoing partnership with your delivery partner once the programme lands: a named contact, an annual allocation of expert requests, and new use cases delivered as they surface. The line organisations forget to budget.
Paid to your AI vendor
AI platform licensing
$155k–$503k/yr
Not a line your delivery partner bills. You pay this to Anthropic, OpenAI or whoever you choose. Priced for the 170 people this programme puts agents in front of, not your whole headcount. Claude Enterprise as a benchmark: base seats ($20/user/month) plus consumption, billed in USD. Licensing the whole organisation instead would be $912k–$3m/yr, a separate decision, and a bigger one.
Your own capacityReal cost, but not a cash line. Keep it separate in your business case.
Not a cash spend
Your internal team
~373–533 person-days
283398 from your core team, plus 90135 across the champion cohort. Worth $256k–$366k of team time at a £500/day fully-loaded rate. Nobody invoices you for it, but it is a real opportunity cost and a real resourcing ask, mostly your programme manager, the process owners in Define, an IT and security lead, and your champion cohort. See the second view of the timeline above for when each is needed.
Year one, all in
$1.4m–$2.3m
Three one-off pillars + first year of managed services + licensing · ex-VAT
Year two onwards
$238k–$756k/yr
Managed services + licensing · the number to check you can sustain
Indicative payback
10 months – 2.4 years
All three levers · Deloitte puts typical enterprise AI payback at 2–4 years · 15 months at best, longer on cautious assumptions on hard savings alone

Five sections follow, one for each part of that picture: who you need internally, how the programme runs, what you pay your AI platform, what recurs after go-live, and whether it pays back.

1

Your Internal Team

The people you need in place, all but one drawn from existing headcount

A programme needs protected time from people who already have full-time jobs. With one exception these are not new hires, but they are not side-of-desk either. Across the roles below the commitment is roughly ~373533 person-days, worth $256k–$366k of team time at a £500/day fully-loaded rate. Nobody invoices you for it. Budget for it anyway. It is the line that derails delivery.

The second view of the timeline above shows the same roles against the eight phases, if you want to see when each is needed.

RoleCommitmentWhat they need to do
Executive sponsor
CEO, COO or CIO
~2 days/month
Full programme duration
Champion the programme visibly at board level, approve the announcement and key communications, and remove blockers. When the CEO mentions AI in an all-hands and references their own use of it, that does more for adoption than fifty training sessions. Needed most at Excite and at every Checkpoint gate, where the go/no-go decision is theirs.
Programme manager
Your counterpart to the delivery partner's engagement manager
12–13 months full-timeDay-to-day coordination, internal navigation and scheduling across the business. The most underestimated requirement on this list: it needs protected time, not a side-of-desk commitment. Heaviest through Define and Design, then steady through the build.
Process owners
One, one per domain you transform
~1 day/week during DefineThe people who own the process being reimagined, they know where the work goes, which exceptions matter, and what 'good' looks like. Their availability in Define is the single biggest predictor of whether the agents you build are the right ones. Not necessarily department heads; whoever owns the process end to end.
IT / security lead
Infrastructure and compliance
3–4 weeks intensive, then ~1 day/monthAI platform configuration (Claude Enterprise, ChatGPT Enterprise or equivalent), SSO, security controls and data access. Concentrated in Define and Design, with a security review before the Checkpoint gate, deliberately before build spend is committed, not after. Light-touch from Build onwards.
Champions
est. 50–100 people
1 day training, then ~2 hrs/week
Heaviest at Deploy
Not teachers, but facilitators. Their job is to notice when a colleague is struggling with a task an agent can handle, guide them to the right setup, and follow up two days later. Recruit on curiosity and influence; technical skill is not a prerequisite. This is the group that decides whether what you built gets used.
AI builder
est. 1–2 people
Full-time from Design onwardsThe one role on this list that is a genuine addition rather than redirected time. Designs human-and-agent workflows, builds the evals that tell you whether an agent is working, and owns the agents after handover. Standard HR job libraries do not have this role yet. You will likely define it from scratch or second a technically curious operator from inside the business.
50100

Champions, your most important internal asset

5–10% of your workforce, recruited from within. Not teachers, but facilitators. Their job is to notice when a colleague is struggling with a task an agent can handle, guide them to the right setup, and follow up two days later. Training takes a day; ongoing commitment is roughly two hours a week. Recruit on curiosity and influence, not technical skill. This is the group that decides whether what you built gets used, which is why change enablement starts in week one rather than after go-live.

Consider in-tool AI assistance to extend your champion network

A champion network is the engine of adoption, but it has a "who do I ask?" bottleneck. One fix is to layer AI assistance inside the tools people already use (Slack, Microsoft Teams, Google Chat, or a platform's own integrations) so day-to-day questions get answered in-channel rather than queueing for a person. That lightens the load on champions and extends their span. Worth asking each delivery partner you shortlist what they offer here.

1–2
AI Builders
Plan to hire: from Design onwards

AI Builder, the emerging internal role

As your programme matures from basic adoption into agent-powered workflows, you need someone who sits between IT and the business. Not a developer. Not a trainer. Someone who can translate business processes into the structured context agents need to do real work, and redesign workflows for a world where humans and agents share tasks.

  • Connect agents to your data: set up secure integrations across legacy and modern systems so agents have the context they need
  • Design human + agent workflows: map where agents take over, where humans step in, and who is accountable for what
  • Manage access controls and monitoring: ensure agents operate within the right entitlements, and that you can see and audit what they do
  • Build evals: create the tests that tell you whether your agents are doing what you intended
  • Keep up with the architecture: the agent landscape is changing faster than any other area of enterprise technology. Someone needs to own this as their full-time job.

This role doesn't exist yet in standard HR job libraries. You will likely need to define it from scratch, or second a technically curious operator from within your business.

2

How the Programme Runs

What happens when, who is needed, and how much of it lands in year one

Programme duration
8–16 months
2 solution waves, 2 agents built at a time
Cost per agent
$134k–$173k
Assumes a typical integration load. Integration complexity is what moves this most, and that is what discovery settles.
Live by end of year one
6 agents
The rest follows as the programme continues

Programme timeline

Eight phases. Excite and Define happen once; everything from the Checkpoint gate onwards repeats per solution, which is why a focused first programme runs in months and a broad one runs in years.

Delivery partnerJointYour team
ExciteDefine /domDesign /waveCheckpoint /waveRefine /agentBuild /agentTest /agentDeploy /agent
Phase
Excite + Define
Delivery: 2 solution waves
Gates
Governance
Use case pipeline
Technical & architecture
Change enablement
Communications
Agents, by solution wave
Wave 1 · agent 1
Wave 1 · agent 2
Wave 1 · agent 3
Wave 1 · agent 4
Wave 2 · agent 5
Wave 2 · agent 6
Wave 2 · agent 7
Wave 2 · agent 8
beyond year one

Time runs in months. Bar height shows how heavy the work is that month, not just whether the workstream is engaged. Governance is a thin constant thread that spikes at every gate, change enablement starts light in week one and peaks at every deploy. That change enablement runs from the start is deliberate, and it is the single biggest difference between a programme that gets used and one that goes live and dies. Dashed agents land after year one. Fewer domains, not a faster build, is the lever that brings them forward.

External programme team

The roles included in the programme cost. The team scales to your organisation size.

RoleCommitmentFocus
Programme Director
Full-time for programme durationOwns the executive sponsor relationship, AI Operating Model, and programme-level decisions. In early-stage programmes, the most senior person on the account.
AI Platform Specialists (1–2)
1–2 specialists, full-time from Design onwardsEach specialist owns a cluster of departments, building their AI projects, prompt templates, agents and skills, and custom workflows on your chosen platform. Scales with the number of departments in scope.
Technical Lead
3 weeks intensive through Define and Design, then advisoryAI platform configuration (Claude Enterprise, ChatGPT Enterprise, or equivalent), SSO, security controls, and data integrations. Heaviest through Define and Design, with a security review before the Checkpoint gate, and mostly done once the infrastructure is live.
Change Lead
Full-timeOwns adoption: programme communications, department head briefings, resistance management, and monitoring uptake across the rollout.
Enablement Leads (2)
2 leads, full-time from Design onwardsDesign and deliver foundation training; train, certify, and quality-control subcontracted facilitators. Set the quality bar for all delivery.
Domain Coaches
2–3 days/month per executive4–6 coaches, each matched by functional domain, a former CFO for the CFO, former CHRO for the CHRO. White-glove one-to-one coaching, not generic AI training.
Facilitators
3–5 people, full-time during rolloutDeliver foundation training sessions. Certified by Enablement Leads before delivery. Scaled to the rollout schedule.
3

AI Platform Licensing

What you'll pay your AI platform provider, separate from professional services

⚠ Indicative figures only, using Claude Enterprise pricing as a benchmarkMajor enterprise AI vendors (Anthropic, OpenAI, Google) do not publish full enterprise rate cards. The figures below use Claude Enterprise pricing as a concrete benchmark because the two-part model (base seat + consumption) is broadly representative of how ChatGPT Enterprise and Google Workspace AI are also priced. Treat them as planning anchors, not committed budget, get a real quote from each platform you shortlist.

Enterprise AI platforms broadly use a two-part billing model: a base seat charge for all users, plus consumption-based usage on top. The more your people use the platform, the higher the usage charges, but unlike traditional SaaS, you're not paying for unused seats at a premium rate. Self-serve plans are billed in USD; sales-assisted plans can accommodate other currencies. Figures here convert with the currency selector above for reference, but the source-of-truth pricing remains USD.

Cost by user type

Enterprise AI plans use a single unified seat type that covers chat, document collaboration, and developer tooling. The split below is a planning view of how usage varies by user behaviour (what you’ll spend in practice), not a description of any one vendor’s billing model. Pricing shown uses Claude Enterprise as the benchmark; ChatGPT Enterprise and Google Workspace AI fall in a comparable range.

Who the programme reaches
The other 165 people in scope are priced as everyday users.

These are starting points, not a claim about your business. People in scope comes from a generic staffing mix, which is wrong for anyone whose shape differs, so if you know the real numbers type them in. Nothing is capped.

User typePeopleSeatConsumptionAll-in monthlyPer year
Everyday user
Writing, research, summarising, analysis
165$20–$30$30–$70$50–$100$59k–$139k
Power user
Heavy daily use, long documents, sustained analysis
4$20–$30$400–$700$420–$730$19k–$34k
Agentic developer
Claude Code, agent building, agentic coding tools
1$20–$30$1k–$3k$1k–$3k$12k–$30k
Agent runtime
Your deployed agents running, billed per agent, not per person
8 agents$250–$3k$250–$3k$24k–$240k
Your total170 licensed$76–$247blended per user$155k–$503k

For your organisation

Minimum commitment (seats only)
$41k/yr
Base seat charge for all licensed users at $20–$30/month. This is your floor, consumption and agent runtime are on top.
Typical all-in estimate
$155k–$503k/yr
Seats, consumption for all three user types, and runtime for your 8 agents. That is $76–$247 per licensed user per month, blended.

Enterprise pricing is negotiated directly with each platform vendor, none publish full public rate cards. Shortlist your platforms (Anthropic, OpenAI, Google) and get a quote from each before committing budget.

4

Ongoing Managed Services

The recurring cost line that gets underestimated, or missed from the budget entirely

The change programme builds the foundation. What comes next is less predictable but no less real: every department will surface ideas they want to act on, new starters will need onboarding into AI workflows, and the technology itself will keep evolving. Your internal IT team will be focused on what it should be, keeping infrastructure running, managing internal systems, maintaining the operational baseline. That work doesn't stop, and it leaves little bandwidth for ongoing AI development at the frontier.

Budget for ongoing external AI support. The specific shape will depend on your pace and ambition, but the organisations that sustain AI impact treat it as a capability to keep investing in, not a project to close.

What this covers

Ongoing AI support is delivered through a quota of structured service requests. Different providers call these “Expert Requests”, “Service Credits”, or retained-hours packages. Common examples include:

  • Writing a prompt or template for a specific recurring task
  • Updating an agent after a process, policy or pricing change
  • Installing a connector to a system your agents need to reach
  • Building a skill or automated workflow for one team’s process
  • Writing an eval that tells you whether an agent is still doing its job
  • A short training session as a new department comes on board
  • Coaching your internal AI lead or champion network

Bigger pieces of work, a department audit or a roadmap refresh, draw on multiple requests rather than one, or are scoped as a project in their own right.

What your programme implies

Expert requests a year
Company-wide work: policies, job descriptions, governance, IT and security25
Care of your 8 agents: prompt updates, integrations, monitoring, re-evaluation after a model change1632
Departmental asks and power-user support across 1,000 people1020
Planning range5177 a year

At $1.6k–$3.3k per request, depending on who delivers it. A request is one of the pieces of work listed above, delivered and done. The thousandth takes the same work as the first, so expect a flat rate rather than a volume discount.

Ongoing, per year
$84k–$253k
5177 requests a year at $1.6k–$3.3k each. Budget it as a standing line rather than a project: the estate needs looking after for as long as it is running, and the demand grows as you add agents.

These ranges come from observed request volumes rather than a published rate card. Of the three lines, the per-head term carries the most uncertainty, because departmental demand depends on how far the agents reach into daily work, so treat the upper end for a large organisation as a planning band rather than a forecast.

5

The Business Case

Where hard returns come from, and how to measure them

The organisations seeing real AI returns are tracking growth they couldn't have achieved otherwise, not just hours saved. According to Wharton's 2025 enterprise AI research, 46% of enterprises now formally track AI profitability. The three value levers below are where those returns show up: growth first, then cost.

Lever 1

Better output, per process

Levers 2 and 3 below are efficiency: the same work, done for less. This one is different, and larger. A well-placed agent changes what a process produces: a higher win rate, a lower cost to serve, a better yield. That is why it matters which processes you pick, and it is the part of the business case only you can fill in.

Each domain starts at a different percentage, because a point of improvement on a full top line and a point on a narrow cost base are not the same claim. Revenue-side domains start low, anchored to the capacity uplift the research below supports; cost-side domains start higher, because the base is smaller and a pound of cost removed is a pound of profit. Organisations leaning into this report EBITDA improvements of 20% and up. Move the sliders and watch where your own assumptions land against that.

Market & sellrevenue side · 8 agents · $1.2m$1.6m of build cost
Shows up as:Revenue per repCost of customer acquisitionWin rateSales cycle length
$6m gross a year$1.2m to the bottom line, at 20% marginPays back its share of the build in under a year
$1.2ma year to the bottom line, across 1 domain
from $6m gross= a 3% lift in EBITDA (ambition is 20%+)

Every number in this lever is your assumption, not our estimate. We have no way of knowing what a process is worth inside your business. Use it to test whether the programme is worth doing at all, and to compare one domain against another.

Lever 2
Headcount avoidance
Roles you don't hire because AI absorbs the work. Growing organisations add 3–5% more headcount annually to keep pace. AI lets you absorb part of that growth without the hire, a hard saving that shows up directly in your cost base.
Illustrative annual value
$719k–$1.9m
Assumes 0.75–2.0% headcount avoidance · $96k loaded cost per role avoided (salary + NI + benefits) · range reflects growing organisations absorbing AI capacity into avoided hires, not pure layoffs
Lever 3
Vendor, agency and SaaS consolidation
External spend that AI makes redundant, in two forms. First, agency and outsourced knowledge work: content, research, translation, document processing, analysis. Second, and increasingly significant: SaaS tools you don't renew, or don't buy at all, because your AI platform can do the job. Wharton data shows enterprises are funding AI budgets by cutting outside services, up 7 percentage points year-on-year. The SaaS consolidation pattern is earlier, but already visible.
Illustrative annual saving
$90k–$377k
Assumes $301–$753/employee in current outsourced knowledge work · 30–50% reduction · SaaS consolidation additive
Payback, all three levers
10 months – 2.4 years
Cumulative net cash flow turning positive, with the ongoing run rate recurring. The faster end assumes high adoption and the upper estimate on both savings levers, the slower end assumes the reverse.
Hard savings only
15 months
at best, longer on cautious assumptions
Levers 2 and 3 alone, ignoring what the agents produce. The conservative anchor for a CFO review, if the case only works with Lever 1, say so.
Time to cash positiven = 20
<1 yr1–22–3>3

You: 10 months – 2.4 years

17 of 20 successful transformations were cash positive inside two years. The faster ones focused on a few high-value domains; the slower ones went enterprise-wide.

Cash-on-cash returnn = 20
<1x1–2x2–3x>3x

You: 1.3×–3.0×

Annual incremental EBITDA divided by one-time investment. The average across the twenty was 3×, so £1 invested returned £3 of EBITDA a year.

Distributions from Rewired (Lamarre, Smaje & Zemmel, 2nd ed., 2026), Exhibit 2.1. Two-thirds of that sample delivered most of their impact from three business domains or fewer, which is the same finding behind the focus guidance higher up this page. Your own figures assume 50% adoption realisation in year one.

These levers work independently, some organisations see all three, others focus on one. We recommend building your business case around the lever most material to your strategy, not the largest projected total. The illustrative figures above use conservative assumptions, your actual returns depend on how ambitiously you deploy and how well you sustain adoption.

Sources: Wharton Human-AI Research & GBK Collective, Accountable Acceleration, October 2025 (budget benchmarks and ROI measurement). Deloitte EMEIA, AI ROI: The Paradox of Rising Investment and Elusive Returns, August–September 2025 (n=1,854, 14 EU/ME countries, payback benchmarks). Google Cloud, ROI of AI 2025 (n=3,466, 24 countries, first use-case ROI). OpenAI, State of Enterprise AI 2025 (productivity gains).

Appendix

Go deeper: research and benchmarks

Every figure in this calculator is anchored to public research from Deloitte, Google Cloud, OpenAI, Wharton, EY, Barclays and others, grouped by what they support so you can verify any number before approving spend.

What this is

A planning tool, not a proposal

The figures in this tool are indicative ranges based on typical programme parameters for organisations of your size. We haven't spoken to you, we haven't scoped your requirements, and nothing here constitutes a quote, from any delivery partner or AI platform vendor.

AI programme costs vary significantly based on your specific situation, what's already in place, and how ambitiously you want to move. The numbers here are a starting point for an internal conversation, not a number to put in a budget.

How to use it

Share it with your buying group (CEO, CFO, CIO) before anyone has formed a strong view. It's most useful when it opens up the conversation, not closes it.

  • Frame all six budget lines before vendor meetings
  • Align your internal group on realistic scale and investment shape
  • Prepare the questions that matter before scoping begins
  • Identify which value lever is most material to your business case

When you're ready for actual numbers, speak to our team ↗

Take this to your team

Download a clean version of this estimate to share internally, or copy a link that opens this exact configuration for your buying group. No forms. No follow-up unless you want it.

Speak to our team

The share link saves your selections so colleagues open the calculator with the same configuration you've set.

Indicative planning estimate only, not a quote from any vendor. Professional services figures exclusive of VAT. AI platform licensing benchmarked against Claude Enterprise pricing in USD; ChatGPT Enterprise and Google Workspace AI sit in a comparable range. Updated figures and live calculator at kowalah.com/resources/ai-program-calculator

APQC Process Classification Framework (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. The PCF is intended to facilitate organizational improvement through process management and benchmarking, regardless of industry, size, or geography. To download the full PCF or industry-specific versions of the PCF, as well as associated measures and benchmarking, please visit www.apqc.org/pcf.

Process domains follow the APQC Process Classification Framework (PCF) Cross-Industry, version 8.0. Programme shape and phase durations informed by Rewired (Lamarre, Smaje & Zemmel, 2nd ed., 2026) alongside Kowalah's own delivery methodology.

Indicative estimates based on typical programme parameters. Actual costs and timelines vary based on detailed scoping, organisational complexity, and your choice of AI platform and delivery partner. Professional services figures are exclusive of VAT. AI platform licensing figures use Claude Enterprise pricing as a benchmark (base seats plus consumption, billed in USD); ChatGPT Enterprise and Google Workspace AI sit in a comparable range, so get a real quote from each platform you shortlist. Managed services figures assume a post-programme ongoing retainer. Business case figures (Section 5) are illustrative ranges based on stated assumptions, not projections. Budget benchmarks referenced from: Wharton Human-AI Research & GBK Collective, Accountable Acceleration, October 2025.