Useful, under your own key
AI in a CRM, under your own key
Every CRM has an AI tab now and most of it is a summarise button nobody presses. A small number of uses are genuinely worth building, and the difference between the two lists is whether the output saves a person a specific job.
What is worth building
Five uses pay for themselves in businesses we have built for. Each one replaces a specific chore rather than adding a capability.
- Turning messy inbound into a structured record. An email, a voicemail transcript or a web form written in prose becomes a record with a name, a service, an address and an urgency, ready for a human to confirm. This is the single highest value use, because it happens hundreds of times a month.
- Drafting the reply, never sending it. A first draft in your own tone, with the record's context, that a person edits and sends. It saves two minutes each time and the person stays responsible for what goes out.
- Summarising a long history before a call, so somebody picking up an account after a colleague has the thread in four lines rather than reading fifty notes.
- Categorising and routing, deciding which team, which priority and which template, on rules too fuzzy to write by hand but obvious to a reader.
- Extracting from documents. Supplier invoices, signed forms, specifications and reports turned into fields on a record, which removes an entire data entry job in some businesses.
The common thread: every one of these turns unstructured input into structured data, or structured data into a shorter read. Those are the two things language models are reliably good at.
What is theatre
Said plainly, because it is quoted often and rarely used twice.
Predictive lead scores from a general model. A number with no explanation, derived from patterns in somebody else's data, which sales people ignore within a quarter. Scoring that works is derived from your own closed deals and split into fit and intent, which is arithmetic rather than AI.
Chatbots that answer customers unsupervised. In a business where a wrong answer costs money or trust, the review effort exceeds the saving. Assisting your own staff is a different matter and is genuinely useful.
Auto generated forecasts. A model guessing at close dates on top of stages that are already unreliable produces confident nonsense. Fix the stage definitions first and the forecast becomes arithmetic.
Sentiment analysis on emails. Demos beautifully, changes nobody's behaviour. The people reading the emails already know how the customer sounds.
Your own key, and what that changes
We build AI features against your own API key with the provider you choose, rather than reselling capacity through us. Three consequences, all in your favour.
| Bundled AI in a product | Your own key | |
|---|---|---|
| Cost | Per user per month, whether used or not | Per use, typically 5 to 40 a month in total |
| Model choice | Whatever the vendor picked | Yours, and changeable in a config value |
| Data handling | Their terms, sometimes including training | Your account, your agreement with the provider |
| Turning it off | Usually bundled into a tier | One switch, and the cost goes to zero |
| New models | When the vendor gets round to it | The week they ship |
The cost line surprises people most. Real usage across a twenty person business, summarising and drafting several hundred times a month, generally lands in tens of dollars rather than hundreds, because the text involved is short. Paying twenty dollars per user per month for the same thing is the part that does not survive inspection.
The rules we build to
Four constraints go into every AI feature we ship, and they are in the scope document rather than assumed.
A person approves anything that leaves the building. Drafts, never sends. The moment a model can email a customer unreviewed, the failure mode stops being a wasted minute and starts being a wrong commitment in writing.
Extraction is confirmed, not trusted. A record created from a model's reading of an email is shown to a human with the source beside it. This costs five seconds and prevents the class of error that is invisible until it is expensive.
Sensitive data does not go to a general model unless the client has decided it may, in writing, with the provider's terms attached. Health notes, case detail and financial records are named explicitly in the scope, and features that would need them are either built differently or not built.
Everything works without it. The AI layer is assistance on top of a system that functions when the key is removed, the provider is down or the feature is switched off. A CRM that stops working because an external model is unavailable is not a CRM. More on how we scope this in CRM automation.
AI powered CRM vs traditional CRM
The benefits of an AI-powered CRM over a traditional CRM are easiest to see side by side.
| Traditional CRM | AI powered CRM | |
|---|---|---|
| Data entry | Typed by the user | Drafted from email, calls and forms, confirmed by the user |
| Follow up | Reminders the user sets | Suggested next step from the record and history |
| Search | Filters and saved views | Questions in plain language, answered from your data |
| Reporting | Built dashboards | Dashboards plus written summaries of what changed |
| Risk | Stale data | Wrong suggestions if the data underneath is poor |
The benefit of AI in a CRM is real but narrower than the marketing: it removes typing and makes good data easier to use. It does not fix a CRM nobody updates. The CRM marketing trend for 2026 is every vendor adding an assistant on top of the same records, so the quality of those records is still what decides the result.
What the 2026 AI CRM market actually offers
The 2026 AI CRM market falls into three groups. The market leaders among the large suites have added assistants priced as an extra per seat. AI first startups sell a lighter CRM built around an assistant, often with limited reporting and integrations. And custom builds, like ours, add AI features to a CRM you own, calling a model under your own API key so usage is billed at cost.
Which is right depends less on the AI than on the CRM underneath. If a suite already fits your process, its assistant is a reasonable add on. If it does not, paying more per seat for AI makes a poor fit more expensive, not better.
What we build
How AI is built into the system
Assistance where it removes a named chore, under your key, with a person in the loop.
Structured records from prose
Emails, transcripts and free text forms turned into fields, shown to a person with the source beside them for one click confirmation.
Drafts in your own tone
Replies and follow ups written from the record's context, edited and sent by a person, with the draft stored on the record.
Summaries before a call
A long history reduced to four lines, generated on demand rather than nightly, so it costs nothing when nobody asks.
Your key, your switch
The provider and model live in a config value, usage is metered and visible, and turning it off leaves a system that still works.
Questions we get asked
Is AI in a CRM actually useful?
Five uses genuinely pay: turning messy inbound into structured records, drafting replies a person sends, summarising a long history before a call, routing on fuzzy rules, and extracting fields from documents. Each replaces a specific chore. Predictive scores, unsupervised chatbots and sentiment analysis rarely change anybody's behaviour.
What does AI in a custom CRM cost?
Typically five to forty dollars a month in total, billed by your provider on usage, because the text involved is short. That is for a twenty person business summarising and drafting several hundred times a month. A per user per month AI tier costs many times more for the same work.
Whose API key is used?
Yours, with the provider you choose. The model name lives in a configuration value, so you can change provider without a rebuild, your usage is billed to your own account under your own agreement, and turning the feature off takes the cost to zero rather than leaving it inside a subscription tier.
Will our data be used to train a model?
Not under the business terms of the major providers, which exclude API traffic from training by default. Because the key is yours, that agreement is between you and them rather than mediated by us. Sensitive categories are named in the scope and either handled differently or excluded entirely.
What happens if the AI is wrong?
A person catches it, because nothing leaves the building unreviewed and every extracted record is shown with its source beside it. The system also works fully without the feature, so an outage at the provider slows a few tasks rather than stopping the business. Both rules are written into the scope.
Tell us how your business works. We will quote the CRM that fits it.
A 30 minute written brief from you, a fixed price and a two page scope from us within two business days. No call required, no obligation.