AI YOU CAN AUDIT
CRM AI with citations, and a human gate
Four AI attribute modes on any object: summarise, cited web research, prompt, and classify. Every proposed value lands in a review queue that shows exactly what accepting writes and the sources it rests on. Nothing touches a record until a human says yes.
AI attributes
Four modes. One column type. Any object.
An AI attribute is a column you add the way you add a text or formula field, on Companies, People, Deals, or any custom object. It runs in one of four modes: summarise the record from what the CRM already holds, research on the web and return an answer with the citations it read, answer a prompt of your own per record, or classify the record into one of the options you define. That last one turns a fuzzy judgement ("is this account strategic?") into a filterable select.
- All four modes propose values; none of them writes without a human accept
- Accepted values carry the AI provenance tag, so a filter can always separate them from what people typed
- Classify outputs behave like any select: filter, group, and report on them
The human gate
Nothing writes until you say so.
Every AI-proposed value routes through one review queue. A review that shows less than the decision needs is a rubber stamp, so the queue shows everything.
1. The AI proposes
An attribute mode runs and its candidate value enters the queue. Nothing has touched the record yet.
2. You see the whole decision
The queue shows exactly what accepting will write, next to the citations the answer rests on. Open the sources before you decide.
3. Accept writes. Reject remembers.
Accepting writes the value with an AI provenance tag. Rejecting writes nothing, and the rejected candidate is kept on the job as provenance, so "why is this field empty" always has an answer.
The same gate governs enrichment. Ishara's enrichment chain (a free 743k-brand directory first, then external providers) runs under write policies, and one of them is suggest-into-review: provider suggestions route into this same queue, so one screen controls everything a machine wants to write to your data.
Beyond attributes
Intelligence that states its reasons.
The rule across the product is the same: no oracle numbers. Every score, brief, and insight says how it got there.
Deal scores, risk, and a coach
Open deals get a score with the reason attached and a risk read beside it. The coach panel on the deal workspace goes further: a diagnosis, concrete next steps each with its why and timeframe, and the one question to answer next.
Grounded prep briefs
Before an investor meeting, a one-minute brief assembled only from corpus facts plus the warm paths in your own CRM. It does not speculate about the person; it organises what is actually known. Part of investor matching.
Multi-provider, one policy
AI work routes across multiple model providers under a single workspace policy. Admins set what is allowed once, and every AI feature inherits it, so provider choice is a workspace decision, not a per-feature surprise.
What "AI you can audit" means
Most CRM AI is oracular: a number appears, and you are asked to trust it. Ishara takes the opposite position. Every AI surface has to answer the question "how do you know": match scores carry an explicit reason for every ranking, research answers carry the citations they rest on, call insights point at the utterance they came from, deal scores state their reason, and the review queue keeps even rejected candidates as provenance. That is the practical meaning of CRM AI with citations: not a disclaimer, a paper trail.
Worked example: one research attribute, end to end
You add a research attribute, "Latest funding round", to Companies. For one record it proposes "Series B" with the two web sources it read. The reviewer opens both, agrees, and accepts. The value lands on the record tagged AI, and the job that produced it keeps its citations. A year later, someone asks where that value came from: the answer is two clicks, not a shrug. Had the reviewer rejected it, the record would show nothing, and the rejected candidate would still sit on the job explaining why.
What Ishara's AI will not do
It does not act on its own. It does not send email, move deals between stages, or edit records unprompted, and we do not claim otherwise. AI in Ishara proposes; people dispose. When you want things to happen automatically, use automations: versioned, deterministic rules whose runs are inspectable and whose failures land in a dead-letter queue. Keeping AI on the propose side of that line is why an auditor, an investor, or your own future self can trust what is in the CRM.
The fastest way to test the philosophy is to open the app, add an AI attribute to a sample object, and watch the review queue show you exactly what it wants to write, and why.
Frequently asked questions
What are the four AI attribute modes?
Summarise the record from what the CRM already holds; research on the web, returning an answer with citations; answer a custom prompt per record; and classify the record into one of the options you define. An AI attribute is a normal column on any object, including custom ones, and accepted values are provenance-tagged as AI.
Does Ishara's AI ever write to a record on its own?
No. Every AI-proposed value goes through the review queue, which shows exactly what accepting will write and the citations behind it. Accepting writes the value with an AI provenance tag; rejecting writes nothing and keeps the rejected candidate as provenance. Enrichment providers can be set to suggest-into-review, which routes their suggestions through the same queue.
Where do the citations come from?
Research-mode answers return the web sources they read, and the review queue displays them next to the proposed value so you can open them before deciding. The job that produced an accepted value keeps its citations. Elsewhere the same rule applies in kind: call insights cite the utterance they came from, and match scores carry an explicit reason.
Which AI model providers does Ishara use?
Ishara routes AI work across multiple providers rather than binding to one. Which providers and features are allowed is a workspace policy that admins set once and every AI surface inherits, so governance lives in one place instead of per feature.
Stop working the market blind.
The investors, the tenders, and the signals are in Ishara from day one. You bring the deals.
Free to start · 32,822 investors on day one · Qatar tenders live, more markets on the way