Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence Stack Buying Map: Which Vendor Covers What

A buyer's map of the layers in a company AI agent stack and the vendor categories that cover each one, so gaps show up before you sign.

Be Human typographic cover with the title: Superintelligence Stack Buying Map: Which Vendor Covers What
Conceptual illustration created with AI · Be Human

The AI agent stack vendors you need are not one product. A company agent stack has seven layers, from the role and its instructions down to monitoring and change control, and each layer is usually covered by a different category of vendor, if it is covered at all. Before you sign anything, map each layer to a named owner and a vendor category, then check where two layers meet, because most gaps sit at those seams.

Key takeaways

  • Buy against layers, not logos. A vendor that covers one layer well still leaves the others to someone.
  • Permissions and approval thresholds are the layer most often left without an owner.
  • Memory and tools are where company data leaves your control, so their boundaries belong in the contract, not only in the demo.
  • Observability and evaluation decide whether you can tell the agent is working after the first month, not just after the first demo.
  • A managed operations provider can cover the seams, but only if its scope is written down layer by layer.

Why buy by layer instead of by product

Most vendor conversations start with a demonstration of one capable tool. The demonstration shows the model answering well, which is the easiest layer to see. What a business actually runs is the combination: an agent that knows its role, reads the right records, acts in the right systems, asks a person before it spends money or sends a customer message, and leaves a trail someone can audit. No single vendor category covers that combination.

Mapping layers first shows you which parts you already have, which you would have to build, and which must be bought as a service with clear limits. It also keeps the purchase honest. A platform that promises the whole stack is often covering two or three layers in depth and the rest in outline, and the outline is where the surprises arrive.

The seven layers and the vendor categories behind them

Use this list as the rows of your map. Each row needs an owner, whether that is a person on your side, an existing tool, a new vendor or a managed provider.

  • Role and instructions: the job the agent performs, what it must never do, and who approves changes to the wording. Your own documented process covers this. No vendor can write your policy for you.
  • Memory and knowledge: the approved records and documents the agent may retrieve. Covered by document stores, search indexes and retrieval tooling. The design of what is retrieved, and from which version, matters more than the brand of storage.
  • Model access: the language model that reasons and writes. Covered by model providers. The buyer questions here are data handling, retention and version control, not benchmark claims.
  • Channels: where requests arrive and replies leave, such as email, a helpdesk queue, a chat tool or a messaging app. Covered by the channel platform and its API permissions.
  • Tools and integrations: the actions the agent can take in your systems, such as reading an invoice or creating a draft order. Covered by integration platforms, direct APIs and connector standards such as Model Context Protocol.
  • Permissions and approvals: which identity the agent uses, which scopes it holds, and which actions require a person. Covered by identity and access management, plus your own approval rules. This is the layer most often unowned.
  • Observability, evaluation and operations: logs, traces, recurring tests, incident response and change control. Covered by monitoring tools, evaluation frameworks and managed operations. The OpenTelemetry observability primer describes the basic vocabulary of logs, traces and metrics that a vendor should be able to explain in plain terms.

For a fuller walk through each layer, the seven-layer stack explanation gives a plainer description you can share with colleagues.

Where the gaps hide: the seams between layers

Gaps rarely sit inside one layer. They appear where two layers meet and each assumes the other is responsible. Four seams cause most of the trouble in practice.

The first is memory to channels. The agent answers a customer from a copy of a price list that was replaced last month, because the retrieval index refreshes on a schedule nobody owns.

The second is tools to permissions. The integration works because it uses a broad service account, so the agent can do far more than its job requires, and nobody reviewed the scope when the connector was added.

The third is evaluation to production. Tests pass in a staging environment, but no one samples live outputs, so a drift in answer quality shows up only when a customer complains.

The fourth is channels to approvals. The channel tool has its own send permission, so a draft that was supposed to wait for review goes out anyway.

Security guidance such as the OWASP authorization cheat sheet stresses checking access where the action happens, not only at the front door. The NIST AI Risk Management Framework similarly treats governance, mapping and measurement as ongoing functions rather than one-time approvals. A regular permission review is the simplest way to catch the second and fourth seams before they cause harm.

A step-by-step plan to map your own stack

This plan takes a leadership team a working session or two. It does not require a technical architect to start, though one should review the result.

  • Write every job the agent will do as one sentence, and name the person who approves changes to that job.
  • For each job, list the records it reads, the systems it writes to and the channels it uses.
  • Place the seven layers beside each job and mark who or what covers each one today: an internal owner, an existing tool, a new vendor, or nothing.
  • Mark every empty cell as a gap, and every cell with two owners as a conflict to resolve.
  • Check each seam by asking which system decides whether an action is allowed, and whether that decision is logged where someone can read it.
  • For each vendor under consideration, record the layers it owns, the layers it touches but does not own, and the company data it would hold.
  • Choose the first task worth handing to an agent, and keep its scope smaller than the full map so the first live run stays observable.

The output is a single page with rows for layers, columns for jobs and a short list of open gaps. That page becomes the agenda for every vendor call that follows.

Illustrative scenario: a quote follow-up agent

The following scenario is illustrative and does not describe a real client. A services firm wants an agent that drafts follow-up messages for open quotes. The map shows that model access, the helpdesk channel and a CRM connector are all covered by existing vendors. Permissions are not. The connector runs on an administrator token, which means the agent could change prices as well as read quotes.

The gap is found before anything is signed. The team changes the connector to a scoped identity that can read quotes and create drafts only. It adds a rule that every outgoing message waits for a person to approve it. It also schedules a weekly sample of live drafts for review, which covers the evaluation seam. The vendor purchase is smaller and clearer as a result, and the first task is easy to explain to the sales team.

How Be Human handles the layers in a managed engagement

Be Human designs company AI infrastructure around one useful role. A first engagement defines that role, the memory and tools it needs, the approval boundaries and one observable success criterion, so every layer in the map has a named owner from the start. Managed operations then cover the seams described above: monitoring, recurring evaluations, integration maintenance, permission review and incident response. The work is remote and available in English and Portuguese. Teams that want to compare this with building in-house or buying a platform can read build, buy or managed before deciding.

If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Do I need one vendor for the whole AI agent stack? Usually not. The layers are covered by different categories of tools, and a single platform that claims all of them tends to be deep in some layers and thin in others. The practical goal is to know which vendor owns each layer and where the handoffs are written down.

Which layer should I check first before signing? Start with permissions and approvals, because it is the layer most often left without an owner and it determines what the agent can do to your data and customers. Once you know the identity, scopes and approval rules, the other layers are easier to judge. Memory and tools come next, since they decide what the agent can see and change.

Can a managed operations provider replace internal ownership? No. A provider can run monitoring, evaluations, integration upkeep and permission reviews, but someone inside the company still has to own the business outcome and approve changes to the agent's role. Written scope and a named internal decision owner are what make managed work accountable.

How do I compare vendors that claim to cover the same layer? Ask each one to show the controls in that layer, not only the feature list. Look for how access is scoped, what data is retained and where, how changes are logged, and how a bad output is detected and reversed. The answers to those four questions usually separate a real capability from a promising demonstration.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →