AI Strategy
When Your Software Vendor Sells You AI: A Four-Question Test
Every established software vendor is arriving at the same conference slide this year: AI, embedded in the product, available soon on a new tier. If you run a business on an ERP, a CRM, an accounting package and a handful of specialist tools, you will be offered AI several times over in the next twelve months.
Before you sign, it is worth understanding why the offers are coming, and having a test ready.
Why the offers are coming
The per-seat subscription model was built on the cost of building software. That cost has collapsed. Small teams with AI coding tools now ship in weeks what once occupied a vendor’s roadmap for a year, and an ordinary assistant subscription covers much of what specialist features used to bill for separately. Vendors know this. Their response is to embed AI, price it as a tier, and keep customers inside the product. It is a rational response to a business model under threat, and it is not in their interest to explain that context to you.
What to keep
The vendors still own something valuable: they are the system of record. Your ERP is where stock and orders authoritatively live. Your accounting package is the truth about money. Your CRM holds the customer list. Those roles are earned through years of stability and edge cases, and replacing a system of record because AI exists is an expensive way to acquire risk.
Keep them, as systems of record. The question is only where the intelligence sits.
The four-question test
Run this on any AI add-on you are offered.
| Question | What a good answer looks like | What a poor answer means |
|---|---|---|
| 1. Can we get this data out? | An API on our current tier, or scheduled exports we control | Data is effectively trapped; any AI value is the vendor’s to price |
| 2. Does the add-on see beyond this system? | It can read our other systems too | It answers only questions inside one silo, which are rarely the questions that matter |
| 3. Which model, and who upgrades it? | Named model, changeable by us | Whatever the vendor licensed, changed when they get to it |
| 4. If we switch vendor or model in two years, what survives? | Our data layer, our logic, our prompts, all in our accounts | Everything built inside the add-on is lost |
If the answers are mostly in the right-hand column, you do not have a decision to make. You have a bill to negotiate.
The pattern we recommend
Four parts, each modest.
- Inventory the systems and their doors. Which have an API, on which plan, and which offer scheduled exports. Ask the account manager early; this answer shapes everything.
- Build one neutral layer you own. Each system feeds plain, consistent tables into a small database in your own cloud account. Reports, automations and your team’s AI assistants all read from it. The vendor is unaware of, and irrelevant to, which model you use.
- Put governance on the layer, not in the vendor. Who may see what, which figure is the official number for each headline metric, what an automation may do without a person’s approval, and a log of every run. This is where the “is it being used the way we are happy with” question gets answered.
- Stay model agnostic by construction. The AI model is one adapter among several. If a different provider leads next year, the swap is one file, not a migration.
The two scenarios this protects you from
The seat pitch. Your ERP vendor prices AI per user, per month. With your data already flowing to your own layer, the evaluation is a single comparison: what does the add-on deliver that your existing assistants, reading the same tables, do not? Usually nothing. Occasionally one specialised capability, which you buy on its own.
The model shift. In two years the best model is from a different company. If your logic lives in your layer, you switch in an afternoon. If it lives inside three vendors’ AI features, there is nothing to switch.
Where this stops
None of this argues for ripping out software that works, building your own ERP, or refusing every vendor feature. Some embedded AI will be genuinely good at one narrow job inside one system, and buying it on its merits is fine. The principle is only that the vendor should hold the record, and you should hold the intelligence.
If a vendor has put an AI tier in front of you and you would like a second opinion before the renewal, send us the proposal. We will run the four questions with you.
If this sparked something, let's talk.
No pitch, no pressure — just a conversation about what you're working on.
Let's talkRelated posts
AI Strategy
Inverting the Approval Pyramid: A Governance Model for Bottom-Up AI
Traditional organisations route every new initiative up to a few executives for approval. AI makes experiments nearly free, so that gate becomes the constraint. Here's an operating model: a sandbox anyone can use, a register, a promotion path, AI-prepared approval briefs, and a smaller, sharper role for the top.
AI Strategy
Did AI Just Kill BI? What to Stop Paying For, and Where to Spend Instead
Traditional business intelligence spent most of its budget on integration: ETL, data wrangling and merging systems. AI has collapsed that cost. Here's what leaders should stop funding, what deserves the money instead, and the questions to put to your BI vendor.
AI Implementation
The AI-Native Stack for a Small Business or Not-for-Profit
Before you subscribe to a website builder, an email tool, a scheduler and a CRM, consider four layers instead: one Google Workspace, a website you own, Claude connected to your data, and Canva. Here's the architecture, what it replaces, and where it stops.