The measure of a good AI project is not how clever it is. It is how many hours it hands back.

Four things AI is genuinely doing for charity teams inside Salesforce this year, what each one costs to build, and the two we now talk clients out of.

The AI work that survives contact with a real charity team is unglamorous: funder reporting, enquiry triage, duplicate matching and supporter self-service. Each one takes four to ten weeks to build on an org that is already tidy, and costs between £6,000 and £20,000. The projects that fail are the ones aimed at judgement rather than admin.

The short answer
  1. Funder and grant reporting
  2. Enquiry and helpline triage
  3. Duplicate matching and data hygiene
  4. Supporter and member self-service

Written and checked by our nonprofit practice team.

Where the week actually goes

Every discovery we run opens with the same request, and it has nothing to do with technology. We ask the team to talk us through a normal Tuesday.

The answers rhyme. A fundraiser spends the morning assembling figures for a report due Friday. A helpline manager retypes the same intake details into a second system because the two do not speak. Somebody in finance is merging duplicate supporter records before the board pack goes out. None of it is the work anyone was hired to do, and all of it expands to fill whatever time is left over.

AI is being sold hard into exactly this gap. Some of that pitch is warranted. A good deal of it is not. Over the past eighteen months we have built, tested and in several cases abandoned AI features for charities across the UK and Ireland, and a fairly clear pattern has emerged about which ones survive contact with a real team on a real Tuesday.

These are the four that consistently do, followed by the two we now advise clients against.

1. Funder and grant reporting

Grant reporting is where the most experienced person on a small team loses the most days. The information a funder wants almost always exists somewhere in Salesforce, but it sits across three objects, in fields that were named by somebody who has since left, and it has to be reassembled by hand into whatever format that particular trust has asked for.

This is the ideal shape for AI, because the judgement has already been made. The numbers are the numbers. What is expensive is the retrieval and the formatting, and that is precisely what a well scoped agent is good at.

The data is nearly always already there. The cost is in getting it out in the shape a particular funder wants.

What this looks like in practice

A grant-making foundation in the south east: 34 reports a year

The foundation makes small grants to community organisations and reports back to its own funders four times a year, plus an annual review. Pulling the figures for a single report took most of two days, and it always fell to the same person.

We built a reporting agent against a tidied data model. It answers plain questions about the grant portfolio, drafts the narrative sections from the case notes already on record, and writes the finished figures back to the grant record so the next report starts from something rather than nothing.

The result: reporting time down from roughly two days to under three hours per report, and a grants manager who no longer books annual leave around the reporting calendar.

A children’s hospice in West Yorkshire: restricted funds without the spreadsheet

Restricted fund tracking had been living in a spreadsheet that three people maintained and nobody trusted. The finance team could tell you the total raised. They could not quickly tell you how much of a particular appeal remained unspent.

Moving that tracking into Salesforce was the unglamorous half of the job. The AI layer on top simply answers the question in plain English, which turned out to be the part that got people using it.

The result: restricted balances available on demand rather than at month end, and the spreadsheet retired after eleven years.

2. Enquiry and helpline triage

Charities running a helpline or an advice line face a version of the same problem every contact centre has, with one important difference. The consequence of getting it wrong is not a poor satisfaction score. It is somebody in difficulty waiting longer than they should.

This makes triage the use case that needs the most care and returns the most value. The rule we work to is straightforward: AI routes, humans respond. An agent can read an enquiry, classify it, pull the relevant history and put a summary in front of the right person. It should not be answering the person in need directly.

Triage is the highest value use case and the one that needs the most careful boundaries.

What this looks like in practice

An advocacy charity: intake in one pass instead of three

Enquiries arrived by email, web form and phone, and every route had its own way of capturing the same information. Casework staff were rekeying details that had already been given once, which is both wasteful and, for the person contacting them, quietly demoralising.

A triage agent now reads inbound enquiries, creates or matches the case record, flags anything with a safeguarding marker for immediate human review, and drafts a summary for the caseworker. Nothing is sent to the enquirer without a person reading it first.

The result: intake handling down from around 40 minutes to 12, with the safeguarding flag rule reviewed monthly by the casework lead rather than left to run unattended.

An advice line in Ireland: the same person, recognised across three channels

Somebody might book a training place, ring the advice line and give monthly, and appear in the system three separate times. Quarterly board reporting involved a fortnight of manual deduplication.

Matching at the point of contact, with a review queue for anything close but not certain, collapsed those records. The board pack now takes an afternoon.

The result: 31% of contact records identified as duplicates and merged, and quarterly reporting reduced from two weeks to one afternoon.

“Kontechs were so different to any other Salesforce partner as they just understand charities inside and out.”

Louise Quall, Head of Development, Ronald McDonald House Charities Ireland

3. Duplicate matching and data hygiene

This is the least exciting item on the list and the one we would build first. Every other use case rests on it. An agent answering questions against a database with three versions of the same supporter will answer confidently and wrongly, which is worse than not answering at all.

Traditional deduplication tools match on exact strings. They miss the interesting cases: a married name, a work address on one record and a home address on the other, a typo in a postcode. Fuzzy matching handles those, provided a person still reviews anything below a confidence threshold you have chosen deliberately.

18% of contact records in a typical charity Salesforce org are duplicates by the time it reaches its fifth year.

Kontechs pre-discovery audits, 34 organisations in the UK and Ireland, 2024 to 2026

The figure surprises people, and then they check. It is rarely negligence. It is five years of imports, three staff changes and a web form that never checked whether an email address already existed.

4. Supporter and member self-service

Most of what a supporter contacts you about is administrative. Changing a direct debit date, updating an address, finding out whether a Gift Aid declaration is on file, downloading a receipt. These arrive by email and get answered by hand, often by somebody whose job is meant to be relationship building.

A portal with a well built assistant handles the routine cases and passes anything unusual to a person, with the context already attached. The important design decision is what it is not allowed to do. Ours never process a cancellation without a human, because a cancellation is a conversation, not a transaction.

Routine requests handled without a queue, anything unusual passed to a person with the history attached.

What this looks like in practice

A membership body: 1,200 members, one part-time administrator

Renewals arrived in a fortnight-long wave every spring, and the same twelve questions came with them. A self-service area answered eleven of them and escalated the twelfth.

The result: around 70% of renewal enquiries resolved without staff involvement, and an administrator who spent the renewal window on the members who actually needed talking to.

What we talk people out of

Two requests come up in almost every conversation, and we now advise against both.

Writing fundraising appeals

It works, in the narrow sense that you get text. But your appeals are your voice, and a supporter who has given for fifteen years can tell when that voice changes. The time saved is small and the thing risked is the relationship. Use it for a first draft of a thank you letter if you must, and have a human rewrite it.

Predicting which supporters will lapse

The models are real and they work at scale. Most charities do not have the scale. With 4,000 supporters and six years of patchy history, a lapse prediction is an expensive way to produce a guess, and it tends to be trusted more than it deserves precisely because it came from a system.

We would rather say this at the first meeting than take the money.

What each one costs and how long it takes

Indicative, and assuming your Salesforce org is in reasonable shape. If it is not, the tidying comes first and it is usually the larger number.

Use caseBuild costElapsed timeBuild it first?
Duplicate matching£6,000 to £11,0004 to 6 weeksYes, always
Funder reporting£9,000 to £16,0006 to 8 weeksAfter the data is clean
Enquiry triage£12,000 to £20,0008 to 10 weeksOnly with a safeguarding review
Self-service portal£10,000 to £18,0006 to 10 weeksOnce volumes justify it
Kontechs project ranges, UK and Ireland charity pricing, August 2026

Licence costs sit on top and vary with your Salesforce edition. We put the full figure in writing before you sign anything, which is what the 95% scoping benchmark exists to protect.

None of this works without somebody inside the organisation owning it after we leave.

Where to start

If you are weighing this up, the useful questions are not about the technology. They are about where the hours go.

  • Which single recurring task takes the most senior time in your team, and how many hours a month is it?
  • Does that task involve judgement, or retrieval and formatting? Only the second one is a candidate.
  • Do you know your duplicate rate? If not, that is the first thing to find out, not the last.
  • Who will own the thing once it is live, and is that written into their role?

If the answer to the last one is nobody, build nothing yet. An unowned automation becomes a liability in about eighteen months, and we have been called in to unpick enough of them to say that with some confidence.

Questions we get asked

Do we need Agentforce, or will something cheaper do?

It depends on volume and on where the data sits. For a small charity with a tidy org, flow automation and a well designed report will often deliver most of the benefit at a fraction of the licence cost. We will tell you which one you need.

Is our supporter data safe if we use AI?

It can be, with the right controls. An agent only sees the records that the person it is working for is permitted to see, and Salesforce’s Einstein Trust Layer is designed so your data is not retained by the AI model provider. Special category data, such as health or safeguarding information, needs particular care under UK GDPR: complete a data protection impact assessment, document your lawful basis, and keep a person in the approval path for anything consequential.

Does this work on NPSP or only Nonprofit Cloud?

Agentforce can be added to most Salesforce orgs, including those running NPSP on a supported edition. The prebuilt nonprofit agents and actions are designed for Agentforce Nonprofit (formerly Nonprofit Cloud), so how much you get out of the box depends on which data model you are on. We confirm exactly what is available for your org during pre-discovery.

How much of our staff time will a project take?

Less than most teams expect, but not none. We need a named owner on your side, time from the people who do the work today for discovery workshops and testing, and someone to own the system after go-live. We agree the time commitment with you during scoping, before anything is signed.

What happens if the AI gets something wrong?

Every use case in this article is designed so that a person reviews the output before it reaches a supporter, beneficiary or funder: AI routes, humans respond. Mistakes are caught at review and fixed at source, and anything with a safeguarding marker always goes straight to a human. We also review how each agent is performing after go-live and adjust its instructions where needed.

Can we start small and expand later?

Yes, and we usually recommend it. Duplicate matching is the natural first step, because every other use case depends on clean data. Once that is in place, each further use case builds on the same foundations rather than starting again.

Sources and further reading

  1. Kontechs pre-discovery audit data, 34 organisations, 2024 to 2026
  2. Salesforce, Nonprofit Cloud data model documentation
  3. Information Commissioner’s Office, guidance on AI and data protection
  4. Charity Commission for England and Wales, guidance on managing charity resources