Tooling · 8 min read · Aug 15, 2026

Five layers. One of them decides everything.

Every tool list for AI in business is either a directory of 200 products or a pitch. This is the stack we would assemble for a group running several entities on mixed systems, by layer, with the honest note about which layer actually determines whether any of it works.

Five layers of an operator's AI stack: interface, agents, connectors, model tiers, and the data foundation everything above depends on.
The stack, top to bottom. Most attention goes to the top two layers; most outcomes are decided by the bottom one.

Layer 5 — The data foundation

Listed last in every vendor deck and first here, because nothing above it survives without it. This is reconciled numbers, an entity hierarchy that matches your legal structure, and access control shaped the same way. If your outlets disagree on what "sales" means, no model resolves that — it will produce a confident answer built on the disagreement.

There is no product for this layer. It is the work described in consolidating reporting across POS systems and multi-entity data isolation. Groups that skip it buy the four layers above and conclude that AI does not work for their business. They are half right: it does not work on their data.

Layer 4 — Model tiers

Not a model — tiers, with routing between them. As of August 2026 that means a reasoning tier (Claude Opus 5, GPT-5.6 Sol) for judgement, an everyday tier (GPT-5.6 Terra) for drafting and summarising, and a fast tier (GPT-5.6 Luna, Gemini 3.5 Flash-Lite, Claude Haiku) for high-volume extraction and classification.

Choose at least two vendors and keep the model a configuration value. Four significant releases landed in under two months this year; anything hard-coded will be re-migrated. The routing rule and the pricing are in the 2026 model tiers.

Layer 3 — Connectors

The Model Context Protocol is now the default way to expose your POS, accounting, HR and storage systems to any AI client, and the 28 July 2026 revision made it stateless, cheaper to host and far stricter about authorization. Prefer an MCP server per system over bespoke per-model integrations — the point is to write the connector once and let the model layer change beneath you.

Check that anything you buy authenticates against your existing identity provider. What that change means in practice is covered in MCP went stateless.

Layer 2 — Agents

Platform-maintained agents are worth buying for commodity work: Salesforce Agentforce and the vertical agents embedded in accounting and HR suites now cover invoice handling, compliance checks and data verification competently. Buy those. Build only where the process is specific to your estate.

Whatever you buy or build, scope each agent to a bounded job with a checkable output. The 88% of pilots that never reach production mostly failed this test rather than a technical one.

Layer 1 — Interface

The least interesting layer and the one most often over-built. A manager will not open a new application to check something they already see elsewhere. Deliver into what they already use — the existing dashboard, the ERP screen, a scheduled message — and reserve a dedicated interface for work that genuinely has no home.

If a group has to be trained to open your tool, the tool is competing with their habits. Systems that get used are the ones that appear inside a routine that already exists.

What not to buy yet

  • Anything priced per seat across your whole staff before one process is in production. Adoption is the constraint, not licences.
  • A second reporting layer. If it disagrees with your close reports, you have added an argument rather than an answer.
  • An agent for an irreversible process. Journal posting and payment release come last for good reason.
  • Anything that cannot show you its working. On a regulated process, an unexplainable output is unusable however accurate.

The honest summary

Four of these five layers are procurement. One is engineering on your own data, it takes the longest, nobody demos it, and it decides the outcome. That is the whole reason this list is shorter than the ones with 200 tools in it.

Dealing with this in your own group?

We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.

Questions

Short answers,
in full.

The questions this article gets asked most, answered so each one stands on its own.

Talk to us
What AI tools should a multi-entity business use in 2026?

Think in five layers rather than products: a reconciled, entity-aware data foundation; at least two model vendors with routing between reasoning and fast tiers; MCP servers exposing your POS, accounting and HR systems; bought agents for commodity back-office work and built ones only where the process is specific to you; and delivery into interfaces staff already open. Four layers are procurement decisions; the data foundation is engineering.

Which layer of the AI stack matters most?

The data foundation, and it is the only one with no product to buy. If outlets disagree on what a metric means or the entity hierarchy does not match the legal structure, models will produce confident answers built on that disagreement. Groups that skip this layer typically buy everything above it and conclude AI does not work for their business.

Should we standardise on one AI vendor?

No. Keep the model a configuration value per job and run at least two vendors. Four significant model releases arrived within two months in 2026, and prices and capabilities moved with them. Teams that embedded a single model in application code have re-migrated repeatedly; teams with a routing layer changed a setting.

What should we not buy yet?

Company-wide per-seat licences before a single process is in production, since adoption rather than licensing is the constraint. A second reporting layer that disagrees with your close reports. Agents for irreversible processes such as journal posting or payment release. And anything that cannot show its working, which is unusable on a regulated process however accurate it is.