Every vendor says they have AI. Ask them these four things.
Today any vendor will show you a chat window. The difference between a demonstration and a capability that supports critical operations lies in four questions that almost never appear in requirements matrices — and one of them we can answer with a third party, not with our own word.
Was it assessed by someone external? →
Yes. AI Capability is one of the twelve processes Pink Elephant certified under PinkVERIFY, with formal criteria and on the standard commercial version of the product. Put that question to the other vendors and compare the answers in writing.
Where does it run and where does our data go? →
Generative models run in your own organization's Azure OpenAI tenant. The statistical models — those that classify and predict — are trained locally on your operation's case history. Your information does not leave to train a third party's model.
How do we know it does not invent answers? →
Every answer carries an openable citation to its source, and a validation checks each citation against the whitelist of material actually retrieved. What is not in the source is not asserted. Personal data is redacted before anything reaches the model.
Can we switch it off? →
Yes, capability by capability, through configuration. And if a model is unavailable or untrained, the platform does not stop: it keeps operating with its classic behaviour. Degrading gracefully is a design requirement, not an accident.
What it does, in operational terms
Eight capabilities, inside the process
Artificial intelligence is not a layer bolted on at the end. It sits inside the process, governed by the objective that process must meet, and it leaves the same auditable evidence as any other activity.
Answers before a ticket exists
A conversational portal where the user describes the situation in their own words. It searches across FAQs, procedures, documents, the glossary and documented solutions, and answers with citations you can open and check. It holds a multi-turn conversation, so a follow-up question keeps the thread. Only when knowledge does not resolve it does it move on to creating a case — and then it collects the related data the category needs step by step, with the process's own validations and no forms to design, flags similar open cases and offers to update the one the user already has. If the question needs a specialist, "ask the experts" routes it to one, and the expert's answer joins the knowledge base on its own.
Classifies using your own history →
A statistical classifier trained on the case history of your own installation, retrained every night, so it learns your taxonomy rather than a generic one. Above a configurable confidence threshold it proposes a single category and asks for confirmation; below it, it offers the best candidates. The model suggests, the person decides.
Flags the deadline risk on day one →
A regression model estimates how long a case will take, contrasts it with its service level and calculates the probability of breaching it. Past a configurable threshold, the case shows a risk alert. The prediction is stored with the case, so afterwards you can audit the forecast against what actually happened.
Summarises a long history →
For cases with dozens of comments, reassignments and events, it produces the state of play: where the case stands, what has been done, what comes next. It is built from the recorded history — it summarises what exists rather than inventing continuity. The technician taking over a reassigned case reads seconds instead of fifteen minutes.
Finds what was already solved
One similarity engine on four fronts: documented solutions and similar past cases with the fix that worked; related problems, linking the case to one in a single click; a warning when the solution you are about to document already exists, so the knowledge base does not fill with duplicates; and detection of near-identical open cases before another one is created.
Points at the change that may have caused it →
When an incident appears, it surfaces recent changes that are candidates for the cause, within a window you configure. It is one of the most expensive questions to answer by hand, because it means correlating the incident timeline against the change log — and it is exactly the question an auditor asks.
Agents, under declared limits →
An AI agent is registered with its limits — allowed actions, confidence threshold, human approval gate and daily cap — and every action it proposes passes through that policy and is logged with its reason. An activity whose executor is an agent is handed automatically to the registered agent that holds the tools it needs, and closure keeps the human gate. And Dexon BPM speaks MCP: Microsoft 365 Copilot, Claude or your own agents discover and use the platform’s tools under each user’s own Entra ID identity and permissions, every call audited.
Keeps junk out of the service desk →
Every inbound email is scored for spam, with its probability and its reason. The case is always created — nothing is lost — and newsletters and junk arrive flagged, so technicians spend their time on real requests. Each confirmation a technician gives retrains the model on your own mail.
How it is governed
Three principles your security team will ask about
What happens to the text
Fragments going to the model pass through personal-data redaction first. Every citation the model produces is validated against the whitelist of material actually retrieved, so an assertion with no source behind it does not reach the user.
Sentiment analysis scores end-user messages in the background and surfaces the emotional state of a case, alerting on deterioration. Spelling correction is non-destructive and queries in another language are translated before retrieval — so search works without the user having to learn how to address a machine.
And this one is not a service desk
Will we close the year on target?
For each strategic indicator, the platform projects the remaining periods of the fiscal year. It extrapolates the trend of the compliance percentage — more stable than the raw value when the target changes between periods — and rebuilds each future period against that period's real target, respecting whether more is better or worse for that indicator.
It answers the management question that usually arrives too late: if we carry on like this, do we finish the year meeting the goal? With enough time left to correct course.
Each indicator carries its estimated year-end close with a 95% prediction interval, and the dashboard flags it amber when it is at risk and red when a miss is likely — months before year-end. It projects only periods that have not been measured: where a real measurement exists, it does not invent a projection over it. And the same arithmetic forecasts demand for every service in the catalogue, with its 95% interval and its trend — capacity planning on the real operation.
Verify it in twenty minutes
Do not take our word for it. Test the classifier against your own history.
Take fifty already-closed cases from your operation, run them through the classifier and compare the proposed category against the one your team assigned. It is a short test and it gives you a real number on your data, not an impression of a rehearsed demonstration. Every capability on this page has a test procedure like that one, and we hand them over.
What we commit to
The deflection figure, built from your own operation
It is common for a vendor to commit to a figure — "cuts volume by 30%" — that wins the paper evaluation and is breached in year two, when the penalty regime kicks in.
Instead of a number decided in advance to win the evaluation, we commit to the mechanism: the platform records every deflection and the usefulness of each answer, so the number is built with you from your own operation and is verifiable month by month.