Softcat
Hand the copying and chasing to AI, and keep people in charge
Agentic workflows use AI to plan and carry out several steps across your business systems, within limits you set. People stay responsible, with approval, monitoring and a fallback wherever a mistake would matter.
When to call us
Repetitive work that follows rules
If people spend their days copying between systems and chasing updates, some of that work can go to an agent.
A process crosses several systems, and people spend hours copying, checking and chasing.
The queue keeps growing, but most of the work follows rules with clear exceptions.
Teams have built agents, and nobody can see what they can access or what they did.
You want faster triage and drafting, without AI sending or deciding everything on its own.
A proof of concept saved time, but nobody planned for the exceptions or the upkeep.
Why it matters
Agents are arriving faster than the controls
Most companies plan to use agents, and very few have the governance in place to let them act on their own.
-
21%
of companies planning to deploy agentic AI in the next two years have a mature model for governing it.
Source: Deloitte, State of AI in the Enterprise, January 2026, opens external site
-
Over 40%
of agentic AI projects will be cancelled by the end of 2027, Gartner predicts, because of cost, unclear value or weak risk controls.
Source: Gartner, Agentic AI forecast, June 2025, opens external site
-
5%
of companies have the full set of controls in place to give AI agents real autonomy.
Source: BCG, Applied AI Index 2026, September 2026, opens external site
What we do
-
Map the process first
Every step, system, volume, delay and exception, so we know what we're changing and can measure the difference.
-
Decide what the agent may do
For each step, whether AI can read, draft, recommend or act, and where a person must approve.
-
Build and connect it
Agents and automations connected to your tools, with access limited to what each step needs.
-
Test the awkward cases
Not just whether the final answer is right, but whether each step, tool call and exception is handled.
-
Log, monitor and hand over
A record of every action and approval, a way to stop it, and an owner on your side.
Worth knowing
What agents can and can't do
Three things worth knowing before you hand a process to AI.
Agentic doesn't mean fully autonomous
Most good first agents draft and recommend, and a person approves anything that matters. Autonomy is earned step by step.
Not every process needs an agent
If the rules never change, ordinary automation is cheaper and more predictable. Agents earn their place when the inputs vary.
Review where it counts, not everywhere
A person checking every step makes the agent pointless. Put approval where the consequence or the uncertainty is.
Self-check
Is this process a good first candidate?
Five questions for any process you're thinking of handing to an agent.
- Does it happen often enough to be worth automating?
- Can you write down what a good outcome looks like?
- Would a mistake be caught before it reached a customer?
- Do the systems involved have a way in, like an API or export?
- Is there someone who'd own it once it's running?
Five yeses make a good pilot. A no on the third question means starting with drafts, not actions.
Questions
The things people usually ask
Plain answers about agentic workflows and business process automation, what's involved and what we need from you. Anything else, just ask.
Which tools can agents work with?
Most business systems with an API, including CRMs, ticketing, finance and document tools. Where there's no API, we look at other routes or tell you it isn't a good candidate.
What does it cost to run?
Model usage plus hosting, monitoring and the time people spend on exceptions. We estimate all of it up front, not just the model bill.
Can we switch AI providers later?
Yes, with some work. We keep the prompts and test cases separate from the model so a switch can be tested before it's made.
