Материалы журнала опубликованы на английском языке. Русская версия сайта

AIAI · AGENTS · OPERATIONS

AI Agents for Business: What to Automate First and What to Keep Human

Where AI agents pay off in a business: how to pick the first workflow, how much authority to give an agent, the risks to control, and what a first project costs.

The short answer

Start with a frequent, measurable workflow that has clear inputs, a costly manual bottleneck and a safe point for human review: triaging enquiries, preparing documents, qualifying leads, reconciling records. Give the agent a narrow set of approved tools, log every action and route uncertain cases to a person. Keep irreversible, high-value and legally sensitive decisions human until the evidence says otherwise.

Weighing this for your business? Request a project review

Illustration: AI Agents for Business

Our take

Axiom Forge designs an agent as a new operational role with a written job description: what it may read, what it may do, what it must hand over and who owns its results. Many first projects need something simpler than an agent, and we say so when a rule or a form would do.

Gevorg AntonianFounder & CEO, Axiom Forge
In this article6 sections · 4 min

Find the workflow before the model

The best first candidates combine repetition with judgment. Someone on your team reads an incoming request, looks something up in two systems, decides what kind of case it is and writes a reply. The work is frequent, it follows a pattern, and its quality can be checked.

Poor candidates are rare events, tasks nobody can describe, and decisions where one error causes serious harm. An agent added to a process that nobody owns will fail for the same reasons a new employee would.

Not every automation is an agent

Anthropic’s engineering guidance draws a useful line. In a workflow, models and tools follow steps that the developers defined in advance. In an agent, the model decides its own steps and which tools to use. Its advice is to find the simplest solution possible and to add complexity only when it demonstrably improves the result.

Many business problems are solved by a workflow with one or two model calls in fixed places. It is cheaper, faster and easier to test than an agent, and it can grow into one later.

Decide how much authority the agent has

Authority should be granted in levels and earned with evidence. Write the level down for each kind of action before the build begins.

Four levels of authority for an AI agent
LevelThe agent mayControl
ObserveRead approved recordsSource and purpose are logged
RecommendDraft a decision or a replyA person approves each one
Act with approvalQueue an action in a business systemA named approver confirms
Act within limitsComplete a bounded routine aloneLimits, policy check and a full log

The risks to design for

Language models can be steered by text hidden in the documents and messages they read. This is prompt injection, the first item in OWASP’s list of risks for LLM applications. The same list warns against excessive agency: an agent with more tools and permissions than its job requires.

The defences are architectural. Give the agent the least access that does the job, separate what it reads from what it is allowed to do, require approval for actions that cannot be undone, and keep a record of every step. Frameworks such as the NIST AI Risk Management Framework help to organise this work, and companies that operate in the European Union, or whose AI systems or outputs are used there, should also check their obligations under the EU AI Act. This is general information, not legal advice.

Measure a pilot against a baseline

Before the pilot, record how the work is done today: time per case, error rate, cost and customer response time. Run the agent on real cases with a person reviewing its output, and compare.

Keep the manual path open. If the model, a tool or a data source fails, cases should fall back to people without being lost.

What a first project costs

A pilot for one workflow that uses systems with existing APIs is usually a matter of weeks. Our AI work starts from $6,000 and from two weeks. Integration with older systems, strict data rules and high volumes add to that.

Plan for running costs as well: model usage, monitoring, evaluation and the time of the person who owns the result.

Discuss your project with the founder

Questions leaders ask

What is the difference between an AI agent and a chatbot?

A chatbot answers questions. An agent can also take actions in business systems through approved tools: create a ticket, update a record, prepare a document. That is why an agent needs defined permissions, logging and a human owner.

What should a company automate first?

A frequent, measurable workflow with clear inputs and a safe point for human review, such as triaging enquiries, preparing standard documents or qualifying leads. Avoid rare, high-stakes decisions as a first project.

How much does it cost to build an AI agent?

A pilot for one workflow with existing APIs usually takes weeks, not months. Our AI work starts from $6,000 and from two weeks. Running costs for model usage and monitoring are separate.

Sources

  1. Anthropic — Building effective agentsanthropic.com (opens in a new tab)
  2. OWASP — Top 10 for LLM Applicationsgenai.owasp.org (opens in a new tab)
  3. NIST — AI Risk Management Frameworknist.gov (opens in a new tab)
  4. EUR-Lex — Regulation (EU) 2024/1689 (AI Act)eur-lex.europa.eu (opens in a new tab)

How Axiom Forge can help

Turn the guidance into an accountable product plan.

Axiom Forge connects product direction, UX, design and engineering for aI product development and business process automation. Start with the business outcome, the people who must use the product and the operating constraints behind it.

Founder-led review · NDA on request · No obligation

See services and starting prices

Write to us

Email us — we reply in English, Russian or Armenian.

Our time now:

Analytics preferences

We use Google Analytics and Google Ads to understand which pages help visitors and to measure enquiries. Form contents and personal details are never sent to them.