Discuss your project
AI Solutions

AI agents for business processes

An AI agent does more than answer questions: it reads a request, works out the steps, uses your systems to gather information and prepares or completes the work, asking a person to approve anything that matters. We design agents for well-defined business processes, give them only the access they need and log everything they do.

Discuss your project How the work is delivered

A good fit if

  • Staff move information between email, spreadsheets, your ERP and other systems all day.
  • Requests arrive in many formats and need to be understood before anything can be done.
  • A process has clear rules most of the time, but exceptions need judgment and context.
  • Rule-based automation has reached its limits because inputs vary too much.
  • You want automation that a person supervises, with approvals and a clear audit trail.

Another approach may suit you better if

  • The process is fully predictable with structured inputs. A conventional workflow automation is cheaper and more reliable.
  • You want an agent to make final decisions about people, credit, health or legal matters without human review. We will not build that.
  • The systems involved have no API or supported integration and cannot be given one.

What this service is

AI agents are software that use a large language model to understand a goal, plan the steps and use tools, such as your ERP, CRM, email or document storage, to carry them out. Where a chatbot answers a question, an agent gets work done: it reads an emailed purchase order, checks the customer and stock in your ERP, prepares the sales order and asks a person to approve it.

We build agents for specific, well-defined business processes and treat them like any other system that touches your data: designed, tested, secured, monitored and supported.

Where agents help

  • Order and request intake. Reading orders, quotes and requests that arrive by email or form, checking them against your systems and preparing them for approval.
  • Finance operations. Preparing supplier invoice exceptions, matching remittances to open invoices and drafting follow-ups for review.
  • Operations and supply chain. Checking stock and supplier lead times, flagging delays and preparing purchase suggestions in Prometheus or ERPNext.
  • HR and onboarding. Assembling onboarding and offboarding tasks across Microsoft 365, applications and equipment requests.
  • IT operations. Triaging internal tickets, gathering diagnostic information and preparing routine changes for an engineer to approve.
  • Reporting. Gathering figures from several systems, drafting commentary and flagging anomalies for a person to check.

Designed for control

The risks of agents are real: they can misread instructions, be manipulated by malicious content in an email or document, or take an action nobody intended. Our designs address this directly:

  • Least privilege. Each agent has its own service identity with only the permissions its process needs.
  • People approve what matters. Actions that change records, spend money or contact customers go to a person for approval, at least until results justify otherwise.
  • Guardrails and limits. Inputs are checked for prompt injection, actions are limited by type, volume and value, and uncertain cases are handed to a person.
  • Full audit trail. Every step, tool call and outcome is logged, so you can see what the agent did and why.
  • A pause button. Agents can be stopped immediately, and the process continues manually.

Agents and your existing automation

Agents work best alongside conventional automation. Predictable steps run as fixed rules in your workflow automation platform or ERP, while the agent handles the parts that need reading and judgment. This keeps costs down and makes the whole process easier to test and audit.

Personal information handled by an agent remains subject to privacy and sector laws, and decisions about individuals need particular care. Some states regulate automated decisions about people: Colorado's law on automated decision-making technology for consequential decisions takes effect January 1, 2027, with notice, an explanation of adverse decisions and human review on request, and California's regulations require businesses that use automated decisionmaking technology for significant decisions to comply by January 1, 2027. This is general information, not legal advice. See AI governance and security for how we assess these risks, and our overview of AI laws in the US.

What is included

The exact list is agreed in writing for each project. These are the usual deliverables and the usual boundaries.

Typical deliverables

  • A process map showing each step, its inputs, the systems involved and where judgment is needed.
  • An agent design defining goals, tools, permissions, approval points, limits and fallbacks.
  • Integrations with the systems the agent uses, such as Prometheus or ERPNext, CRM, email, Microsoft 365 and file storage, through their APIs.
  • Least-privilege service accounts and credentials stored securely.
  • Human approval steps for actions that change records, spend money or contact customers.
  • An evaluation set of real cases and results showing accuracy and exception rates.
  • Audit logs of every step and action, with alerts for failures and unusual behavior.
  • A dashboard of volumes, outcomes, exceptions and running costs.
  • Documentation, training and a runbook, including how to pause the agent.

Not included unless agreed separately

  • AI model, platform and cloud usage charges, which you pay to the provider.
  • Changes to third-party systems beyond their supported configuration and APIs.
  • Fully autonomous decisions about individuals without human review.
  • Ongoing support after handover, unless agreed in writing.

What we will need from you

Most delays in this kind of work come from access and decisions, not from the technical build. Knowing these early keeps the project predictable.

  • A process owner who can explain the rules, the exceptions and what good looks like.
  • Representative historical cases, with personal information protected as agreed.
  • API access or integration options for each system the agent will use.
  • Agreement on which actions need human approval and who approves them.
Delivery

How the work is delivered

Each stage ends with something you can review before the next one starts.

  1. Map the process

    Walk through the process with the people who run it, measure volumes and identify where an agent would help and where people must stay involved.

    Output: Process map and agent opportunity.

  2. Design the agent

    Define the agent's goals, tools, permissions, approval points, limits, logging and fallbacks.

    Output: Agent design and risk review.

  3. Build and test

    Build the agent and its integrations, then test it on historical cases and edge cases in a non-production environment.

    Output: Tested agent and evaluation results.

  4. Supervised pilot

    Run the agent on live work with a person approving every action, and track accuracy, exceptions and time saved.

    Output: Pilot results and tuning.

  5. Scale with oversight

    Reduce approvals only where results justify it, extend to more volume, and monitor continuously.

    Output: Live agent, runbook and monitoring.

Testing and handover

  • The agent meets agreed accuracy and exception thresholds on historical cases before going live.
  • Every action is logged with the inputs, reasoning summary, tool calls and outcome.
  • Permissions are limited to what the process needs, and tested.
  • A person can pause the agent instantly, and work continues manually.
  • Spending and volume limits prevent runaway costs or actions.

What affects the cost

We do not publish package prices. Each estimate is based on an agreed scope, in US dollars, with taxes shown separately. These are the things that move the number most:

  • The number of steps, systems and integrations in the process.
  • The variety of inputs and how many exceptions there are.
  • How many actions need approval and how approvals are handled.
  • Security, privacy and audit requirements.
  • Model and cloud usage, which grows with volume.

Questions buyers usually ask

What is the difference between an AI agent and automation we already have?

Traditional automation follows fixed rules on structured inputs. An agent can read unstructured requests, decide which steps apply and use several systems to complete them. We often combine both, with rules for the predictable parts and an agent for the parts that need interpretation.

Can an agent make mistakes?

Yes, which is why we design for it. Agents start with a person approving every action, have limited permissions, log everything and send uncertain cases to a person. Approvals are reduced only where results prove it is safe.

Which platforms do you use?

We build agents with Microsoft Copilot Studio, cloud AI services from Microsoft Azure, Amazon Web Services and Google Cloud, open frameworks and, where data must stay in your environment, open-weight models on servers you control, and connect them to your systems through APIs and open standards such as the Model Context Protocol. We choose based on your systems, security needs and costs.

Which processes make good first agents?

Frequent, well-understood processes with clear outcomes, such as entering orders received by email into your ERP, preparing supplier invoice exceptions for review, assembling onboarding tasks across systems or triaging internal IT requests.

How do you stop an agent from doing something harmful?

Least-privilege access, approval steps for consequential actions, limits on volume and spending, input checks against prompt injection, monitoring and a pause control. We also review the agent's risks with you before it goes live.

Insight

What is agentic AI?

What AI agents are, how they differ from chatbots and automation, where they help in the back office and which controls they need.

Have a process that eats your team's day?

Describe the process, the systems involved and where it gets stuck. We will reply to arrange a conversation about whether an agent, a simpler automation or both would help.

Discuss an AI agent