AI Agent from Cloudify

AI Agent from Cloudify

AI that understands the request

AI that understands the request

AI that understands the request

We design, build, run and maintain workflows that interpret emails, documents and requests

We design, build, run and maintain workflows that interpret emails, documents and requests

Agent running
Shared provider mailbox · 12 weeks live
New email · Outlook · 08:15
Re: transfer request, see attached
From a pension provider, no customer number, one PDF
  1. Reading the new email
  2. Finding the customer
  3. Filing the document
  4. Creating the task
  5. Done, arrival to finish
Done without a person
2,085of 2,489
Handled by your team
40316%
84% done without a person · 2.6 actions per message · 92% of decisions kept

WHAT IT DID

WHAT IT DID

A shared mailbox that files itself

A shared mailbox that files itself

A financial advisory firm receives a continuous stream of correspondence from the companies it works with. Every message used to mean someone reading it, working out what kind of request it was, finding the right customer, updating the record in HubSpot and filing the document against it.

The agent now recognises seven kinds of message and does the matching work for each — or stops, when it cannot identify the customer with certainty. Some categories it completes end to end. Others it prepares and deliberately leaves for a person, because answering on the firm’s behalf was never part of the brief.

A financial advisory firm receives a continuous stream of correspondence from the companies it works with. Every message used to mean someone reading it, working out what kind of request it was, finding the right customer, updating the record in HubSpot and filing the document against it.

The agent now recognises seven kinds of message and does the matching work for each — or stops, when it cannot identify the customer with certainty. Some categories it completes end to end. Others it prepares and deliberately leaves for a person, because answering on the firm’s behalf was never part of the brief.

0
1
2
,
0
1
2
3
4
0
1
2
3
4
5
6
7
8
0
1
2
3
4
5
6
7
8
9

Messages handled in twelve weeks

0
1
2
3
4
5
6
7
8
0
1
2
3
4
%

With the admin done by the agent

0
1
2
3
4
5
6
,
0
1
2
3
4
0
0

Actions carried out in the connected systems

0

Records linked to the wrong customer

Four in five messages now arrive with the work already done. The team reads a decision, not an inbox.

A personThe agent
4–5 minutes per message
About 11 seconds per message
Available Mon–Fri, 8 to 4
Available every day, all day
40 hours a week
168 hours a week
A personThe agent
4–5 minutes per message
About 11 seconds per message
Available Mon–Fri, 8 to 4
Available every day, all day
40 hours a week
168 hours a week

Each message sets off a chain of work in the connected systems rather than a single update — the customer found, the document generated and attached, the record advanced, the task raised, the duplicate check run. Nothing is deleted or moved, and every run is logged.

WHERE IT HELPS

WHERE IT HELPS

Two kinds of work

Two kinds of work

The admin nobody wants to do, and the checking nobody has time to do. Most people come to us about the first. The second is usually worth more.

Today
8:00, coffee first
Mette working through a tall pile of emails
8:009:0010:0011:00
With an agent
8:00, coffee first
Mette relaxed while an agent sorts her inbox
8:009:0010:0011:00
0 min back
Mette’s Monday morning

WHAT IT DID

WHAT IT DID

The agent decides. The integration does the work.

The agent decides. The integration does the work.

The agent works out what something is, checks it against what it can find out, and decides what happens next. A document, a request or a claim arrives, and it is measured against your rules, your records and whatever outside source the case calls for — before anything is written anywhere.

Something arrives
an invoicea claima questiona timesheet
The agent
compares it against what it can find out
Knowledge base
what should be true
Guardrails
what it may not do
System access
read and write
What it checks against
Your systems
Your documents
Outside sources
Does it add up?
It checks out
handled, start to finish
It stops short
admin done, decision left to you
It can’t tell
handed over, groundwork done
Why the plumbing is the hard part

A convincing demo is the beginning, not the finish. Production work also needs reliable access to your systems, the business rules written down, exception handling and ongoing evaluation. Those rules are not in the model. They are in your business.

Illustrative examples

Illustrative examples

Two you'll recognise

Two you'll recognise

These are illustrations of the pattern, not case studies — the delivered example is the pension mailbox above. Every workflow we build gets described the same four ways. The fourth line is the one that matters — it's where we agree, before anything is built, what the agent is not allowed to settle on its own.

Case 01 · Delivery disputes

Delivery disputes, settled while the evidence is still fresh

What arrivesA complaint with no order number, buried in a long thread.
What it looks upThe order, carrier scans and delivery photo, returns policy and claim history.
What it doesEvidence agrees and the value is within your limits: refund or replacement raised.
What makes it stopConflicting evidence or a value above your threshold. A person gets the full trail.
Case 02 · Documentation questions

The questions your documentation already answers, answered

What arrivesA question in the customer’s own words, rarely matching your docs.
What it looks upYour documentation, their plan and version, and earlier tickets.
What it doesDrafts the reply with the right links. You decide whether it sends.
What makes it stopThe docs don’t cover it, or contradict their account.

Controls

Controls

How we keep you in control

How we keep you in control

How far the agent goes is your decision, and it is set per case type rather than per project. The same agent can run end to end on one kind of email and deliberately stop short on another, for as long as you want it to. Underneath that sit the controls, agreed with you in writing before anything is built.

1Level 1 · Suggest

It prepares, you decide

It reads, gathers and recommends. Nothing is written until a person says so.

Agent preparesYou decide
Used for due diligence, credit decisions, anything a regulator expects a named person to decide.
2Level 2 · Do the work

The admin is done, the judgement is yours

It finds the customer, files the document and creates the task, then stops. You spend thirty seconds deciding instead of five minutes looking things up.

Agent does the adminYou decide
About 3 in 10 messages on our live deployment, by design.
Level 3 · Handle it

Start to finish, unless it isn’t sure

Where the rule is settled and a mistake is recoverable, it completes the work. The escalation path stays live underneath.

Agent handles itEscalates if unsure
Just over half of all messages. A further 16% were stopped because it couldn’t identify the customer with certainty.

How we run it

An agentic workflow is delivered as part of a custom integration project — the same four stages as every Cloudify build. The agent changes what is possible in Design and Deliver. It does not change how we work, or who is accountable when it runs.

1

Discover

We map the process as it actually runs, including the exceptions people handle without thinking. Some processes come back as a no.

You get Process map and a go/no-go
Worth building?No: we tell you and stop here
2

Design

We agree what the agent settles, what it never settles and where a person stays in the loop. Written down before anything is built.

You get Written rules, signed off by you
3

Deliver

Built on your real data and run alongside your team until the numbers hold. Handed over with the escalation path live.

You get A live agent, proven in parallel
4

Drive

We watch accuracy and flag rates after launch, and tighten the rules as edge cases surface. They always do.

You get Accuracy and flag-rate tracking
One-timeBuild feeDepends on how complex the process is.
OngoingSubscriptionBased on the systems connected and the volume running through them.
Priced after Discover, not beforeIf it isn’t worth building, Discover is where we say so.

Questions we get asked

Is this a platform we have to learn?
What happens when the agent is wrong?
Which model do you use?
Where does our data go?
How long does a first workflow take?