Almost nobody in this market publishes what AI costs. Vendors quote per conversation, consultants quote per client, and the result is that a finance director in Kampala has no way to know whether the number in front of them is reasonable. So here is a complete stack for a twenty-person Ugandan firm, priced at published list rates, with the date each was checked and the exchange rate stated. Everything here will move. That is a reason to date it, not a reason to keep it private.
Key Takeaways
- A twenty-person firm needs roughly eight paid seats, not twenty. Buying for everyone is the commonest way to overspend in year one.
- Year one comes to about USD 7,100, and steady state to about USD 2,200. The difference is training and pilots, both one-off.
- The two largest lines are the seats and the internal owner’s time, and only one of them appears on a vendor quote.
- Shilling exposure is real: these are dollar subscriptions on a card, and most local cards add a foreign transaction fee on top.
- The cheapest line in the whole stack is training, and it is the one that decides whether the expensive lines produce anything.
The assumptions, stated up front
| Firm size | 20 people |
| Staff who genuinely need a paid seat | 8 |
| Exchange rate used | UGX 3,722 to USD 1, mid-market, 19 August 2026 |
| Prices checked | 21 August 2026, published list rates |
| What you will actually pay | More. Bank and card rates sit above mid-market and foreign transaction fees apply |
Shilling figures below are rounded to the nearest ten thousand and are for budgeting rather than reconciliation.
What a seat costs
These are published list prices as at 21 August 2026. Business tiers matter for more than features: they generally carry stronger written commitments on how your content is retained and whether it may be used to improve the provider’s systems, which is the reason to buy them.
| Plan | Monthly billing | Annual billing | Note |
|---|---|---|---|
| Claude Pro, individual | USD 20 | USD 17 | Annual billed USD 200 up front |
| Claude Team, standard seat | USD 25 | USD 20 | Per seat, per month |
| ChatGPT Business | USD 25 | USD 20 | Two-seat minimum |
| Microsoft 365 Copilot, business | USD 21 per user, per month | An add-on requiring a qualifying Microsoft 365 base subscription. A promotional rate of USD 18 was published as running to 30 September 2026 | |
| Google Workspace Business Standard | around USD 14 per user, per month | Gemini is now folded into the plan rather than sold separately, with base prices raised to match. Published figures differ across sources between roughly USD 14 and USD 18, so check your own quote | |
The stack, priced
| Line | USD, year one | UGX, year one |
|---|---|---|
| 8 business seats at USD 20 per month, billed annually | 1,920 | 7,150,000 |
| Automation tooling, lowest paid tier | 274 | 1,020,000 |
| Team training, two days | 2,400 | 8,930,000 |
| One 30-day pilot, internal time only | 900 | 3,350,000 |
| Internal owner, roughly 4 hours a week for 6 months | 1,600 | 5,960,000 |
| Year one total | 7,094 | 26,410,000 |
| Steady state, year two onward | 2,194 | 8,170,000 |
On the automation line: a workflow tool at the lowest paid cloud tier runs around EUR 20 a month billed annually, roughly USD 274 a year at recent rates, typically with unlimited workflows and users and a monthly execution cap. A self-hosted community edition exists at zero licence cost, which trades the fee for server administration. That is only a saving if the person to administer it already exists.
Read the internal-owner line twice
It is the second largest number in the table and it appears on no vendor quote.
Somebody has to answer questions, fix prompts that stopped working, chase the vendor, keep the tool register current and notice when a team has quietly stopped using something. Without that person, the seats are bought and unused within a quarter. This is the single most common reason a well-funded AI programme stalls in this market: the money was there and the hours were not.
Four hours a week for the first six months is a realistic figure for a firm of this size, tapering afterwards. It is not a new hire. It is an existing person with that time protected and their manager’s agreement, which is a conversation rather than a budget line and is therefore easier to skip.
Why eight seats and not twenty
In every deployment we have seen, the useful ratio has been closer to one seat in three than one in one. The people who benefit are the ones who write, analyse and answer: proposals, reports, correspondence, customer questions, data work. A warehouse supervisor, a driver and a receptionist are not underserved by not having one.
Buying for everyone costs USD 4,800 a year instead of USD 1,920, and the additional twelve seats will show low usage within two months. The saving is real and available and it requires one uncomfortable conversation about who gets a seat. Solve it by function rather than seniority and the conversation is short.
Then review quarterly and take back what is unused. Fifteen minutes a quarter, and in most organisations a quarter of seats go quiet within two months.
The costs that appear on no quote
| Cost | Typical size | Why it lands |
|---|---|---|
| Verification time | 10 to 40 percent of the time saved | Output has to be checked before it goes out. Budgets that treat drafting time as fully recovered are why pilots report savings that never appear in the work |
| Bandwidth | Meaningful, and load-bearing | These tools are useless on a connection that drops mid-session. In much of the region the honest constraint is a stable link, not a licence |
| Power | Varies widely | An outage during a working session costs the session. Where power is unreliable, backup for these machines is part of the AI budget whether or not anyone calls it that |
| Devices | Occasional, lumpy | Older machines and small screens make sustained AI work painful. Some deployments quietly become device refresh projects |
| The failed pilot | USD 800 to 1,500 | The price of finding out. Budget for two pilots and expect one to fail. It is a fraction of the cost of a project nobody can end |
Seven ways to cut the bill honestly
- Buy seats for a third of staff. Largest saving available, by a wide margin.
- Use the assistant inside tools you already pay for. Often removes a subscription line entirely. Costs you capability at the harder end of the work, which is why a handful of dedicated seats usually still make sense alongside it.
- Bill annually where you can commit. Typically 15 to 20 percent. Costs flexibility and requires hard currency up front, which is a real constraint here.
- Ask for the non-profit rate. Many vendors run non-profit programmes at 40 to 60 percent off and Ugandan registration documents are commonly accepted. Rarely asked for, frequently granted.
- Review seats quarterly. Take back the quiet ones.
- Self-host the automation layer. Only a saving if a competent administrator already exists on the payroll.
- Kill the pilot that is not working, on schedule. The largest saving of all and the hardest to book, because it appears as a project that did not happen.
Two false economies
Free consumer accounts for work. They save the licence and cost you the written commitments on retention and reuse that make the tool defensible under a data protection review. Once personal data is involved that is not a theoretical risk.
Skipping the training. It is the cheapest line in the stack and the one that decides whether the expensive lines produce anything. An untrained team uses an AI assistant as a search box for about three weeks and then stops.
The number this article cannot give you
What it is worth. That depends entirely on your work, and any article that tells you a return figure has invented it.
The way to find out is to take one specific repeating task, measure what it costs in hours today, apply a conservative reduction, and check the payback. In year one, on a well-scoped task, 25 to 45 percent is what we typically see, not the 80 percent that appears in vendor material. Include the checking time inside that figure rather than excluding it.
And say out loud what happens to the freed hours: headcount you will not need to add, work you can now take on, or overtime you stop paying. If the answer is none of those, the saving is real for the individual and invisible in the accounts, and it is better to know that before the budget meeting than during it.
Sources
- Claude plan pricing, claude.com, checked 21 August 2026.
- ChatGPT Business published pricing, checked 21 August 2026, reflecting the April 2026 reduction from the previous USD 25 and 30 rates.
- Microsoft 365 Copilot published business and enterprise pricing, checked 21 August 2026, including a promotional rate published as running to 30 September 2026.
- Google Workspace Business Standard pricing following the folding of Gemini into base plans. Published sources differ between roughly USD 14 and USD 18 per user per month, and this is flagged in the text.
- USD to UGX mid-market rate of 3,722, 19 August 2026.
Prices checked 21 August 2026. Nothing here is a quotation and AI pricing has changed several times a year since 2023. This article is reviewed quarterly.