Skip to the comparison

AI agent use cases for travel and expense, compared with what you already run

Two products and the dashboard most programmes already run, side by side across seventeen rows, including three where the dashboard wins.

Help me pick the right oneShow me where they differ
Cogent's cited £1.24m answer and Orchestra's approved review share a PredictX panel, dashboard outside. Figures illustrative.

Every answer cited to source · No new system for travellers · Runs on your existing feeds · ISO 27001:2017 · PCI-DSS · Cyber Essentials · GDPR

You do not have to replace anything to find out where yours stops.

The short answer

AI agent use cases in travel and expense come down to two. One answers a question against your live data in seconds: compliance by traveller, contract utilisation, a policy change modelled across real bookings, duplicate claims caught before reimbursement. The other runs a whole project through to a finished deliverable, under your approval. Instead of reading a chart and writing the conclusion underneath it, you get the conclusion with every figure cited back to source.

The comparison

Two AI agent examples, and the way most programmes work today

IBM defines an AI agent as a system that autonomously performs tasks by designing workflows with the tools available to it (IBM, 2024). Two of the three columns below are agents. The third is the dashboard, BI stack, self-service analytics or spreadsheet most programmes already run.

Your existing reporting almost certainly holds the data. These rows are about what happens between having it and getting the answer. Every column is a complete specification on its own: read down one, or across a row.

Comparing

PredictX

Everything else

Cogent

The one you ask. Put a question to your live travel and expense data and get an answer back in seconds, cited to the record it came from.

Live · Runs on your existing feeds

See what Cogent does on your data

Orchestra

The one you delegate to. An AI agent team scopes a project, runs it under your approval, and hands back a finished deliverable.

Live · Adds to an existing PredictX programme

Meet the Orchestra agent team

A dashboard or BI stack

The dashboard, BI stack, self-service analytics or spreadsheet most programmes already run, and the general-purpose assistant sitting beside it. Described as the work it leaves you doing, not as a claim about any vendor.

Whatever you already have in place

Not a product, and nothing to buy here

Cogent and Orchestra compared with the dashboard, BI stack, self-service analytics or spreadsheet most programmes already run.
AttributeCogentOrchestraA dashboard or BI stack
01
Summary: a question, a project, a chart
Each of the three finishes a different unit of work. That unit is the fastest way to tell them apart, and it is the thing to match against what your programme actually needs.
Unit of work

A question

A project

A chart

The short version
  • Answered in seconds
  • Every figure cited to source
  • 200+ maintained connectors
  • No SQL to write, no BI ticket to raise
  • Exports finance can use
  • One project manager, nine specialists
  • Five approval gates, in order
  • Report, PDF or editable deck
  • Full audit lineage on every re-run
  • Runs on a cadence you set
  • You pick the dashboard view that sits closest to the question
  • You reconcile the feeds yourself first
  • You write the conclusion underneath the chart
  • You raise a ticket for anything it was not built for
  • Already deployed, and your team already knows it
What it covers
  • Travel and expense only
  • 200+ maintained connectors across TMC, card, expense, booking, HR and finance feeds
  • Travel and expense only
  • The same consolidated estate, worked as projects
  • Typically your whole business
  • One stack across finance, HR, sales and supply chain. Cogent and Orchestra stay inside travel and expense, so the dashboard wins this row
02
How the work starts
Who begins the work, and who does the assembling before an answer is possible.
Who starts it

You ask, whenever the question arrives

You commission a project, then approve the plan before anything runs

You open the dashboard you have, or raise a ticket for whatever sits outside it

Who assembles the data

The agent, reading your live feeds directly

The agent team, across nine specialists

Typically done up front, when the BI stack was built. Anything outside that model comes back to you or to an analyst

How many systems you touch

One. Nothing else required

One. Built on Cogent and named for it: Orchestra, powered by Cogent

  • Typically several, and you move between them
  • Dozens of pre-built views across multiple modules, plus a spreadsheet to join them
03
What comes back, and how fast
The artefact you are left holding at the end, and how long it takes to arrive.
What you get

An answer, cited to source, in a thread you can audit

A stakeholder-ready deliverable: report, PDF or editable deck

A dashboard tile or a spreadsheet export. You supply the interpretation

Report or dashboard

A cited answer in a thread, on demand

A report. Editable, exportable, cited to source

  • A dashboard
  • The right tool for monitoring a known metric, and the dashboard wins this row
Time to an answer

Seconds

Hours to days, paused at five approval gates

  • Immediate for a question the dashboard already answers
  • Days to weeks for one it was not built for
04
Approval and control
What runs without a human saying yes, and what each one is allowed to touch.
Human approval gates

None before the answer, though it asks you to clarify when a brief is ambiguous. You interrogate it after: every figure opens to its reasoning, and any thread re-runs on fresh data

Five, in order: define, plan, execute, validate, deliver

Whatever your own review process already is

What it does without your approval

Reads and analyses your data, and returns a cited answer for you to act on

Reads, analyses and drafts. Nothing runs until you approve the plan

Whatever your dashboard and BI stack already do. You act on what you read, and you decide what happens next

05
How you know the numbers are right
Where a figure comes from, and who checks it before anyone acts on it.
Where a figure comes from

Every figure traced to the source query behind it

A provenance trail on every claim in the deliverable

You check the numbers by hand before you present them

Who checks the work

An independent validation agent audits outputs before release

  • Silent correction before the draft
  • An independent data-health audit
  • A provenance trail on every claim

You, or a second analyst, depending on how much the answer matters

06
Getting it running, and what it costs you
Where the dashboard you already run has the advantage.
Availability

Live

Live

Already running in most programmes

Time before anyone can use it

Deploys on your existing feeds. No new system for travellers

Adds to an existing PredictX programme with no new integration work

None. It is already there, your team already knows it, and that is a real advantage. The dashboard wins this row

What your team has to learn

How to ask a question in plain language

How to scope a project and approve a plan

Typically nothing new. The cost is what it takes to get an answer it was not built for

What you are buying

A subscription across the programme, sold by conversation

A step up from Cogent, sold by conversation

Varies widely: per seat, per query, per project, or already inside a platform you pay for

Go deeper
See what Cogent does on your data
Meet the Orchestra agent team

How they relate

The two are not separate systems. Orchestra is powered by Cogent, so the same consolidated travel and expense data sits under both: the agentic AI framework both run on.

You can test the shape of the work before you choose either of them: 50 free AI prompts to run first, no gate and no form. The download appears straight away.

On the control rows

Trust in an agent tracks what it is allowed to touch. The PwC AI Agent Survey of 308 US executives, fielded in April 2025, put trust highest for data analysis at 38% and lowest for financial transactions at 20%. Both products on this page sit in the first group. They read your data, analyse it and return a cited answer, and Orchestra pauses at five approval gates before anything is delivered.

Before procurement asks

Both products inherit the same controls as the framework underneath them: ISO 27001:2017, PCI-DSS, Cyber Essentials and GDPR. They deploy on the feeds you already send, so nothing new is asked of travellers and no data leaves the estate it sits in today.

Where to start

Which one is right for you

Four situations, written as readers describe them rather than as job titles.

“A new question joins the data analyst queue and I wait weeks.”

Start with Cogent.

It closes the gap between the question and the answer, which is the gap that costs most when the question changes every week.

“I have a list of work I never had the team to do.”

Start with Orchestra.

It is built for the analyses a programme rations because the outside version is a consulting engagement that lands weeks after the decision. PredictX puts a comparable external engagement at $50K to $200K, from its own programme experience: that is a market cost for outside work, not a PredictX price. Orchestra is for when the work needs finishing, not answering.

“What we already have mostly works.”

Then keep it, and be specific about where it stops.

The questions worth an agent are the ones where you join the feeds by hand, work out what the chart means, or raise a ticket and wait. If none of those is happening on your programme, you do not need this page.

“Both sound like things I need.”

Then start at the question and add upward.

They stack: Orchestra is powered by Cogent, so moving up the range adds depth rather than replacing what you already run. Nobody has to buy both on day one, and the order is not arbitrary.

In practice

How AI agents start in a finance team: pick one recurring question that already has a deadline, run it on your own consolidated data, and check the answer against the way you produce it today. That is what makes AI agents for finance verifiable rather than impressive. The sequence that works is one recurring question on Cogent first, then one scheduled project on Orchestra once the answers are trusted.

Pick the question where the spend is moving. The 2026 GBTA Business Travel Index, published on 3 August 2026, forecasts global business travel spending at $1.71 trillion this year, a rise of 7.2%, while trips rise 1.3% to 1.84 billion. Cost per trip is what is changing, and that is a different question every month rather than the same view every month.

FAQ

Questions your team will ask

What are the different types of AI agent a business actually uses?
Two, in practice. One you ask, which answers a question against your own data in seconds. One you delegate to, which scopes a project, runs it under your approval and returns a finished deliverable. Most programmes need the first before they need the second.
Is ChatGPT an AI agent?
No, not as most people use it. A general-purpose chatbot answers from what it was trained on and starts blank every time. An AI agent holds a goal, plans the steps, executes them against your live systems and validates the result before you see it. Chat can be how you brief an agent. It is not the agent.
What is the difference between an AI agent and an AI assistant?
Who decides the next step. An assistant waits for the next instruction and helps a person do their own work. An agent is given a goal, decides the sequence itself, dispatches the work across specialists and returns the finished result. Both products on this page decide the sequence. They differ in how big a unit of work they take.
How is an AI agent different from a dashboard?
A dashboard shows pre-modelled figures and leaves the interpretation to you. Every new question becomes a rebuild request. An AI agent takes the question itself, runs the analysis across live data and returns the finished answer with its working shown. Instead of reading a chart and deciding what it means, you read a conclusion and check the evidence behind it.
What can an AI agent do with travel and expense data?
Six things it does concretely: reconcile booking, card and expense records that do not match; score compliance by traveller and cost centre; model a policy change across real bookings before you commit; track contract utilisation against every airline and hotel agreement; flag duplicate and anomalous claims before reimbursement; and consolidate travel emissions by route and entity.
Which AI agent do I need for a travel and expense programme?
Start with the one that matches your unit of work. If the thing you keep waiting for is an answer, that is Cogent. If it is a piece of analysis nobody has capacity to run, that is Orchestra. Most programmes need the first before they need the second.
Do I need both, or will one do?
Most programmes start with Cogent and add Orchestra when the questions turn into projects. Nothing is wasted by starting there, because the two run on the same data and the same connectors, so the step up adds capability rather than replacing what you already have.
Can I use Cogent without Orchestra?
Yes. The dependency runs one way. Cogent is a product in its own right and answers questions on its own. Orchestra is built on Cogent and named for it: Orchestra, powered by Cogent. Programmes already on PredictX add Orchestra with no new integration work, because the connectors and certifications carry over.
How should I compare pricing across AI agent products?
Compare what one unit of work costs, not what one seat or one token costs. Token, credit and per-action models price the machinery. What a programme buys is an answer or a project, so ask what a single one of those costs and whether the price moves when the question gets harder. PredictX publishes no prices.
Why not just use the reporting tools we already have?
For anything your existing views already answer, you should, and the table says so on three rows. The gap opens on the questions they were not built for: the ones where you reconcile the feeds by hand, interpret the chart yourself, or raise a ticket and wait. Those are the questions an agent takes.
How do you run an AI agent pilot that produces reliable signal rather than demo results?
Run it on your own data, on a question you already know the answer to. A demo built on sample data proves the interface works. A pilot that reproduces a figure you spent a week assembling last quarter proves the reasoning works. Give it one recurring question with a real deadline, then check the working, not just the number.
What security questions matter most for agents that can act in your systems?
Four: what it can do without a human approving, what it reads, who may see what it returns, and what record survives afterwards. Cogent reads and analyses your data, then returns a cited answer for you to act on. Orchestra reads, analyses and drafts, and pauses at five approval gates. Both log every agent action, and access is enforced by role.
Will an AI agent replace our data analysts and BI team?
No, it changes what they build. Most analytics capacity goes on servicing requests: the pull, the deck, the one-off question. An agent absorbs that queue, which frees the team for modelling, data engineering and the judgement calls only they can make. The queue disappears. The team does not.

The question gets answered before the meeting, not after it.

Run one of your own questions through both

Bring the question your team could not get to this quarter. We will run it on data shaped like yours, show you which of the two answers it, and say so if the answer is the tooling you already have.

Help me pick the right oneShow me what they share

No new systems for travellers · Your existing feeds · Nothing runs until you approve the plan

4 of the 6 largest global T&E programmes run on PredictX

Most programmes start with one and add the other at its ceiling. Cogent answers the question, and Orchestra finishes the job.