Adjutant AI, AI-driven integration and the Domain Workspace, by The Dutch Data Difference
The questions Splunk teams ask every week, answered by an agent
Eleven of them have an answer you can run today, from alert triage to keeping your CMDB and Splunk in step. Three more are being built right now, and this page says plainly which is which.
Every one of them is an agent running inside your own Splunk deployment, under your own roles, on a model you choose, carrying only the tools, skills and knowledge it was given, and leaving a record of what it did and why.
Available today
Each one names the outcome first, then how it works underneath, then where it stops and which edition it needs.
Which of these alerts actually touches something that matters?
Alerts reach the analyst with the business context that was missing: which service is affected, who owns it, how critical it is and what changed recently. The queue gets worked in order of impact instead of order of arrival.
How it works underneath
- ServiceNow CMDB and CSDM context pulled into Splunk and traversed per alert, with round-trip write-back to ServiceNow
- Triage agents that run the moment a notable fires, with no person in the loop and no write they were not granted
- Splunk Security Essentials content matched against the data you actually have
- Explain any SPL search in the context of your own environment, including why it stopped returning results
Where it stops: it adds context and triage around your existing detections. It does not replace your detection content, and the context is only as good as the asset data behind it, which is what the reconciliation use case below is for. Security workflows are part of Enterprise; ServiceNow access is an optional add-on.
The service is red. Is it the database, or is the feed lying to us?
A reproducible verdict on why an ITSI service is degraded, pointing at the probable downstream cause. When the honest answer is a broken or stale feed rather than a broken system, it says that instead.
How it works underneath
- Deterministic scoring over the real ITSI dependency graph; the model only narrates the result, it never picks the cause
- Same evidence, same ranking, with fingerprints to prove it
- Feed trust checks through TrackMe, Cribl Stream and Splunk Edge Processor health
- Runs hands-off from an ITSI episode action rule and reports back into the episode or by e-mail
- Remediation is drafted as a recommendation only; nothing changes without a person
Where it stops: it needs Splunk ITSI with a modelled service tree. It ranks causes it has evidence for and will not guess beyond that. ITSI workflows are part of Enterprise.
What can our data actually detect, and where are the rings hiding?
An honest grade of nine fraud typologies against the data you have today, before anyone builds a detection. Then investigations that surface what a score threshold never will: many accounts that each look harmless, sharing the same infrastructure.
How it works underneath
- Fraud Data Coverage Audit grading account takeover, synthetic identity, payment, card, mule and ring, cross-channel, claims, insider and AML
- Coverage and joinability graded separately, because present fields are not the same as keys that survive between feeds
- Ring scan inside the Splunk risk index, grouping by shared attribute instead of by account
- A blocked grade names the missing feed and what it takes to onboard it
Where it stops: it is not a real-time authorisation engine. It works around your detection layer, not in the payment path. Fraud analytics is an Enterprise add-on.
Our alert priority depends on asset data that three systems disagree about. Which one is right?
One reconciled view of applications, services, owners and criticality, built from every inventory you already have. Splunk then prioritises security events and service impact on data a person approved, instead of on whichever feed happened to arrive last.
How it works underneath
- Each source is staged and then accepted or rejected as a whole: missing keys, duplicate keys or a read that was cut short never half-feed a decision
- Declared survivorship rules decide which source wins each field, and every conflict is recorded with both values and where each came from
- What cannot be decided safely is quarantined, never guessed, and a quarantined record is never treated as gone
- A record counts as gone only after a grace period of complete observations from a source trusted to say so
- Promotion to the current truth needs two different administrators, and the run proof shows the chain from source row to landed record
- The reconciliation itself calls no AI model at all, and the screen says so rather than leaving the number blank
Where it stops: this is a scheduled reconciliation, not a live asset lookup. It needs Enterprise with the AI-driven integration feature. Any Splunk administrator can act on any organisation’s reconciliation today; there is no per-organisation administrator role yet.
Can that reconciled truth flow back into ServiceNow without the CMDB owner fearing it?
ServiceNow is made to equal the approved view for the tables you choose, and the CMDB owner sees exactly what will change, record by record, before anything is written. Everything the integration did not create is left alone.
How it works underneath
- A preview that is stored and then recomputed at go-live: creates, updates, retirements and deletes, plus everything held back with the reason named per record
- An ownership ledger: only records this integration created are ever changed, and the rest is counted on screen as not ours
- Holds for a cold start, an empty result, a drop below the record floor, or more removals than the cap you set
- Writes go through ServiceNow’s own identification engine, one record at a time, each read back and counted only if it is really there
- Relationships such as application depends on service are checked against the type’s direction, never created twice, and only removed when this integration created them
- A capability test per table and an identification check before the first live run; a run is refused rather than guessed at
Where it stops: authority is per table and per scope you declare, never the whole table. Deletion is opt-in. Retirement uses ServiceNow’s lifecycle API where your instance grants it, otherwise a fallback the preview has to state. Enterprise with the AI-driven integration feature.
Can we move data into Splunk on a schedule and prove every record landed?
Integrations where the mapping is a reviewed contract rather than a prompt. A replay corrects data instead of duplicating it, and a source that changes shape slows a run down rather than silently rewriting it.
How it works underneath
- An AI model orchestrates the run but never touches a record: the rows stay on the server and travel by handle, so ten thousand rows cost what ten would
- Only the fields the mapping lists enter the pipeline, which is your privacy boundary, followed by a closed list of thirteen transformations in order
- Four-eyes promotion, then a verify run against the real destination before a mapping may ever run unattended, with clearance tied to the version
- The watermark advances only after the destination was read back and the landing verified, so nothing is skipped that never arrived
- Lands in KV Store, an event or metric index, or object storage, and only where that agent’s allowlist permits
- Optional stop before every write, so an administrator approves the exact payload that is then sent
Where it stops: delivery is at least once, so keyed destinations absorb a repeat and event indexes record it twice. A read capped at ten thousand rows is partial and holds its watermark deliberately. Scheduled integrations need Enterprise with the AI-driven integration feature.
What changed in the firewall policy, and does it still match what we think we allow?
Snapshots of your Cisco ASA and FMC configuration, normalised into rules, objects and expanded rule tuples, kept in Splunk as searchable history. A change has a before and an after, and a question about whether traffic is permitted becomes a search instead of a reading exercise.
How it works underneath
- The tool chain collects, parses and normalises the configuration; the model does not interpret the rules
- Expanded tuples, so correlation searches can join policy against observed traffic
- Snapshots stored per collection run, so drift between intent and configuration is visible
- Runs unattended on a schedule, with the same history and receipts as any other scheduled work
Where it stops: Cisco ASA and FMC today. It needs the data onboarding capability, which is Professional and above.
Can the nightly checks run without someone babysitting them?
Investigations and reports that run on a schedule, in the right order, as a governed service account. Each run leaves a history entry you can open, and a failed stage triggers its own follow-up.
How it works underneath
- Agents on a cron, with overlap protection so a slow run is never doubled
- Dependency chains with on-success and on-error follow-ons, so each stage starts when the previous one finishes
- Learned facts about your environment, so it stops re-asking; unattended runs use confirmed facts only
- Safe on search head clusters: state lives in replicated storage, not on one node
Where it stops: scheduled jobs need Professional or higher.
I know what I want to see. I just do not want to write the SPL.
Working searches and dashboards from a plain-language question, checked against your own data before you get them. No broken panels discovered the morning after.
How it works underneath
- A dashboard builder that validates and repairs panels before it publishes
- A search validation loop that treats zero results as a question, not a pass
- Sixty-five ready agent templates for recurring Splunk work, twenty-five more in preview, plus your own
- Controls per organisation and business unit on which tools the model may use
Where it stops: it cannot find data you have not onboarded. It will tell you what is missing instead of building an empty dashboard.
Can our corporate AI chat answer Splunk questions without getting the keys to all of Splunk?
Your company’s own AI chat answers Splunk questions for the scenarios you choose, such as a host health check or an incident summary, by running a reviewed agent inside Splunk. It never gets free rein over your indexes, and the same question gets the same kind of answer every time.
How it works underneath
- Tag an agent for MCP and it becomes available to the chat; every agent you did not tag stays invisible
- Published through Splunk’s own MCP Server, next to Splunk’s own tools, or through Adjutant’s standalone MCP endpoint, which offers nothing but your agents; access uses Splunk tokens
- Every call names the organisation and business unit it runs for, and only agents visible to that tenant can be listed or run
- Start and collect: long investigations run in the background without timing out the chat, and every run leaves a history entry
Where it stops: off by default and part of Enterprise. It needs Splunk’s MCP Server 1.2 or later, or the standalone endpoint. In this first release, running agents over MCP requires an admin-level identity, so connect the chat through a dedicated service account.
Why does every answer have to go through the Splunk team first?
A Domain Workspace: one governed surface over a business domain, where a business person asks in their own words and gets a panel that can prove where its number came from. When the data cannot answer, it names the cause instead of drawing an empty chart.
How it works underneath
- One versioned Domain Model declares the domain: its sources, entities, how they join, the questions people really ask and the known-correct answers that prove them
- Every search is compiled on the server from typed parameters and executed by Splunk as the viewer, under their existing roles, so two colleagues see different but equally truthful versions
- Four separate assurances per panel, never collapsed into one green tick, and a question that was never proven can never claim it was
- Evidence snapshots built from references and re-verifiable months later, for the audit question “what did we know when we decided this”
- Recurring findings become a worked case queue with stable work items, claims and notes, instead of a spreadsheet in an e-mail
- Ask something outside the question bank and it refuses and shows what it can answer, rather than running the nearest-looking search
Where it stops: nothing leaves your Splunk instance, and the workspace runs no language model calls until you connect one. Question matching is deterministic against a reviewed bank, so there are no prompts to protect in the first place. A Domain Model is authored, not yet learned from your existing dashboards. The Splunkbase package is the guided preview tour; the running workspace comes through early access. Licensed at Enterprise level, with Adjutant AI Enterprise included.
In active development
Three things are being built right now. They are on this page because customers are shaping them, not because they are finished. Each card says exactly where it stands, and we would rather tell you that than find out later that you expected more.
One thread runs through all three: a Domain Model describes a business domain, and Splunk is the biggest source in it rather than its boundary. The work below is how a workspace reaches the rest of your data, and how people reach the workspace from where they already are.
Can our people ask the Domain Model a question from Microsoft Teams, and be warned there before an SLA breaks?
A Teams app where someone types “where is order ORD-123456” or “why was this payment not delivered” and gets the answer from the Domain Model in your Splunk, with how fresh the data is and a link to the exact view behind it. And when orders are approaching an SLA breach, one card appears in the agreed channel. Not a data feed into Teams: the insight itself, where the conversation already happens.
How it will work underneath
- Every question runs under the asking person’s own Splunk identity and roles, so two colleagues can get different answers to the same question and both are right
- No language model decides an answer: answers come from the Domain Model’s declared questions and reviewed searches, each with a receipt you can trace in Splunk
- Splunk opens no connection to Microsoft or the internet; a small relay that you host is the only component that talks to Microsoft
- Sensitive values are masked before anything reaches Teams, and a channel card carries a verdict and a reference rather than data rows
- The bot only listens when it is addressed, and reads or stores no channel conversation
- A card’s button opens the full Domain Workspace one click away, and as a tab inside the channel where the pilot proves that works
Where it stands: specified end to end, including the Splunk roles, the relay, the identity linking and every network flow, with a brief for each team that has to approve something. The Splunk-side components and the relay are not generally available; a pilot comes first. If your organisation runs Teams and Splunk and wants to be in that pilot, we would like to talk.
Half our data sits in Elasticsearch. Do we really have to copy it into Splunk to use it?
Search Elastic from the Splunk search bar and let the results flow into ordinary SPL: stats, timechart, lookups, joins with Splunk data, dashboards and alerts. Elastic stays the system of record, nothing is re-ingested, and you do not pay a second platform to store the same events. For a SOC that runs Splunk for security and Elastic for application logs, it turns two investigations into one.
How it will work underneath
- New search commands read Elastic in place: what exists, what the schema is, the documents themselves, and aggregations pushed down to Elastic rather than pulled across
- Results keep Elastic’s own field names and types, the ones people already read in Kibana, with the handful that collide with Splunk’s own fields renamed and listed
- Cross-SIEM investigation: one entity or indicator checked on both platforms, with corroboration, contradiction, unique signals and visibility gaps named per platform, and likely duplicates flagged rather than silently merged
- Elastic Security alerts and cases read into Splunk, so one queue carries both and a status change arrives as an update, never as a duplicate alert. One direction only: nothing is written back to Elastic or Kibana
- One access model: which Splunk roles reach which Elastic data is governed in Adjutant, on top of what Elasticsearch itself enforces, with every query traceable from Splunk to Elastic and back
- The assistant sees metadata only by default; raw Elastic values reach a model only where an administrator deliberately raised that setting
- Documentation your teams keep in Elastic, such as runbooks and post-incident reviews, can be designated as knowledge the assistant may consult, off until an administrator names a source, with a link to every source it used
- Elastic also becomes a source for the AI-driven integration above, read by the deterministic plan executor with no model calls at all
Where it stands: the specification is complete, every open decision in it is closed after an independent review, and implementation is under way. None of it is installable yet. The first release covers governed live search, the assistant and the cross-SIEM alert workflow. It will need an Enterprise licence that also carries the Elastic feature, and neither of those on its own is enough. It targets Elasticsearch and Kibana 8.19 and current 9.x minors; 7.17 is out of scope. Work that needs Splunk-indexed or accelerated data, such as tstats over an accelerated data model, stays outside the live-search path. We read the code of the Elastic packages already on Splunkbase before writing any of this: four of the five copy Elastic data into Splunk, the fifth polls into SOAR, and none of them queries Elastic in place from the Splunk search bar. We are looking for design partners who run both platforms, including at least one Splunk Cloud estate.
Our question spans Splunk, Elastic and a system of record. Who answers it?
A Domain Workspace that covers the whole domain rather than the part that happens to be indexed in Splunk. The Domain Model stays the single versioned description of the business: the entities, how they join, the questions and the proof. What changes is how far it can reach for the data behind an answer.
What that gives you
- One model per domain, many audiences: the analyst, the fraud desk, compliance and the executive see one set of numbers at four altitudes, not four dashboards that disagree
- Splunk events joined with reference data from other systems, reached through the same governed path rather than copied in first
- Honest states stay honest across sources: a feed nobody onboarded, an index this viewer cannot read and a genuinely empty answer remain three different, named outcomes
- Reconciled asset, owner and criticality data from the integration work above becomes the context the whole workspace uses
- Every search still executes as the viewer, under the Splunk roles your auditors already accept
Where it stands: what is published on Splunkbase today is the guided preview tour on fictitious data. The running workspace exists, works against a live Splunk instance and ships its own proof, a thirty-check self-test that runs on your estate including negative tests for forged receipts and scope widening, and it is available through early access rather than as a download. What is not built is the part people ask for first, a learner that mines a Domain Model out of your existing dashboards; today a model is authored. Reaching sources beyond Splunk follows the Elastic integration above. This card describes the direction and is the one place on this page where we are telling you where we are going rather than what you can install.
What holds for every use case
It runs inside your estate
Native Splunk apps on your own deployment, using the model endpoint you choose, cloud or on-premises, or none at all. Where another system is involved, it is reached from your side, never the other way round.
Deterministic where it counts
Root cause ranking, data mappings, reconciliation and coverage grades are computed by code. A model may orchestrate and explain; it never transforms a record or invents a number.
Every action on the record
Audit logging, agent telemetry with OCSF export, receipts per run and read-back after every write. Not just AI you can trust. AI you can prove.
New: Splunk Use Case Radar, edition Q4 2026
Where should your Splunk team be right now?
Twenty-one Splunk Enterprise use cases in four rings: what you should already run, what to pilot this quarter, what to learn first and what to stop, such as staying on 9.x past 16 December. Vendor-neutral, with a first check for each one that you can run today.
See one of these on data that looks like yours
Adjutant AI is on Splunkbase as app 8747, with a free Personal edition to try it as one user. For a team, a 30-day Proof of Value with agreed success criteria is the fastest way to evaluate. For anything in active development, a design partner conversation is the way in.
