Prepare and actAgentic AI systems

Agents that prepare the case, and act only where you allow.

Your rules and engines settle the routine; the referrals, exceptions, and missing evidence land on your people. We build agents that gather the evidence, prepare the file, and act only inside the limits you write.

Start with a Decision BlueprintTalk to the engineer

FIG. 1 · Written limit · credit referral agentIllustrativeIn force

The agent prepares the file. A person decides every one.

Investigate
Reads the statements, bureau report, and e-invoicesAllowed
Recommend
Drafts the file, with every source linkedAllowed
Decide
Every file goes to the credit managerStays with a person
Act
None: it never approves, changes terms, or contacts the customerStays with a person
Written limit
None for the agent; above the credit manager's authority, the next level decides
Stops when
Two sources disagree, or evidence is missing

Every step is recorded.

When it fits, and when you don't need us

Use an agent only where the work needs one.

  • The next step depends on what is found

    Which source to check and what to open next change from case to case, and your people open four systems to assemble one file.

  • Routine cases could clear inside your policy

    Cases that meet every written threshold wait in the same queue as the hard ones.

  • You don't need us

    The route never changes

    If every case follows the same steps, a workflow or a rule is enough, and easier to audit.

Not sure which kind fits your decision? The Decision Blueprint answers that first.

How we build it

The rules it works under are written outside the model.

We build the agent's harness first, the written rules and controls around it: what it may read, where it runs, when it stops, and who decides. The model inside can change; the limits stay where you wrote them.

FIG. 2 · Harness · credit referral agentIllustrative

Harness

Tools and access

Read-only by default, under its own login.

  • Statementsread-only
  • Bureau reportread-only
  • E-invoicesread-only
  • Change termsnot granted

Review agent

Gathers the evidence and writes the file; cannot send it.

  • SandboxRuns isolated; nothing it touches leaks out.
  • GuardrailsInput and output checks; defense against instructions hidden in documents (prompt injection).
  • MemoryWhat it keeps on a case, and for how long.
  • Retrieval-augmented generation (RAG), agentic RAGReads your sources at answer time and cites each one.

Stop conditions

Missing evidence, conflicting sources, a limit reached.

Written limit

Written by you; the agent holds none on credit.

Credit manager decides every file

Human in or on the loop

Who approves, who watches, who can stop it.

  • Evals before go-liveTested on your past cases, hard ones included.
  • TracingEvery step, source, and action recorded.
  • Kill switchOne switch stops it.
In plain words: the agent reads your systems but cannot change them, stops when evidence is missing or sources disagree, and hands every file to a named person. Every step is recorded, and one switch stops it.

Inside the limits your policy writes, the system acts; everything else goes to a named person.

The agent never grants itself a tool, a permission, or an approval.

Workflow, or autonomous within written limits

You set the limit, and move it only when the results earn it.

  • InvestigateGathers the evidence.Allowed
  • RecommendProposes a decision, with reasons.Allowed
  • DecideChooses the outcome on routine cases.Stays with a person
  • ActTakes the approved step.Stays with a person
  • A named person decides past the limit

Decision supportThe agent investigates and recommends; a person decides every case.

How the work is arranged

The simplest arrangement that holds wins.

For your technical team: seven arrangements, each shown in a use case

What you receive

An agent you can audit, running where your systems are.

FIG. 3 · HandoverIllustrative

  • agent/The agent, its harness, and its code
  • limits/The written limits and the permission map
  • evals/Tests built from your cases
  • traces/The record of every step, for audit
  • monitoring/Monitoring, the kill switch, and rollback
  • runbooks/How your team runs it, with training

Where it runs

Inside your environment, next to the systems it reads: your own servers or your cloud account.

It starts from the models, rules, and engines you already run; no new model is required.

How your data, model weights, and access are handled: Security & working with us

Desks that use it

Desks where an agent prepares the case.

  • ModelsSLMsAgentic AI

    SME credit referrals

    TodayReferred files wait while analysts rebuild the firm from four systems and last year's committee notes.

    With the systemOne file with the rating, its reasons, and the authority level.

    Credit manager decides

  • ModelsSLMsAgentic AI

    Early warning on the loan book

    TodayThe borrower is current; other banks are already cutting its limits, and you see its quarter only at renewal.

    With the systemA file, not a spreadsheet: signals checked in parallel and gathered in one place.

    Watch-list committee decides

  • ModelsAgentic AI

    Behavioral score and limit management

    TodayLimits are set once at application; good customers outgrow them and quietly deteriorating ones keep full, undrawn lines.

    With the systemLimit actions queued with their evidence, inside the committee's matrix.

    Underwriter decides

  • ModelsSLMsAgentic AI

    Insurance claims fast-track

    TodaySimple claims wait in the same queue as hard ones, and customers wait with them.

    With the systemSimple claims take the fast track through your own process; the system never declines a claim.

    Claims adjuster decides

  • ModelsSLMsAgentic AI

    The lease file

    TodayA dealer sends a proforma invoice; you must know if the lessee is good, the price is right, and the machine is not already financed elsewhere.

    With the systemThe lessee, the price, and other financing checked in one file.

    Credit committee decides

  • SLMsAgentic AI

    Customer complaint 8D

    TodayContainment is due tomorrow, and the evidence is in five systems.

    With the systemEvidence from five systems in one 8D draft.

    Quality engineer decides

See all 6 desks

It works on its own, and stacks with:

  • Machine-learning models Gives the agent a score or forecast to act on, with its reasons.
  • Small language models (SLMs) The model the agent reasons with: reads what it gathers in your field's terms and calls your tools inside the agent's written limits.
  • Integration · the base: Under every kind: we build and integrate, or integrate what you already run. What it connects to

Field note: Boring patterns that keep multi-agent systems up

Equalize the gap.

Bring the work your rules cannot finish.

Tell us where the referrals pile up, which systems they touch, and who signs. The Decision Blueprint shows whether an agent is needed at all.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/agentic-ai · Printed copy; the website is the current version.