Attio

Marketed as an AI native platform, and the billing model shows the AI is a layer rather than the substance. Enrichment, the assistant and the research agents all draw on a metered credit pool, and the vendor states which workflow operations consume credits and which do not: triggers, branching logic, delays and lookups are free, while record writes, external calls and AI blocks are charged.

A capability the vendor meters separately from the subscription is by construction detachable from it. Strip every credit consuming operation and a complete relationship platform remains, with its data model, records, pipelines, email and calendar synchronisation and permissions intact, which is what the reviews describe teams choosing it for.

The category editorial anticipates exactly this, noting that every platform of this kind now ships an assistant and almost none fails the removal test. The AI here is better integrated than most, appearing as first class record attributes rather than a side panel, and it still is not the product. Ask which AI features are included in the tier quoted and what they consume.

A named administrative control over autonomous consumption, at a level of granularity almost nothing else in this index publishes. Credits are split into two pools: a per seat allowance for individual assistant queries and a shared workspace pool powering workflow automation, AI attributes and research agents.

Administrators can configure whether a user who exhausts their personal allowance may fall back to the shared pool, which is a specific, per workspace setting that prevents one heavy user draining the resource everyone else depends on. That is a real oversight mechanism rather than a policy statement.

Alongside it the free versus charged distinction in the workflow engine makes agent activity visible in the bill, since an AI block costs and a lookup does not, so a workflow's autonomous surface is legible from its cost. Held below the top band because these are spend and scope controls rather than behavioural ones: nothing published states what a research agent may do within a run, what sources it may reach, or whether its output is marked as inferred when it lands in a record beside verified fields. Ask what bounds a research agent per run and whether its output is distinguishable from entered data.

Cost transparency stands in for model transparency and does not reach it. The vendor publishes a clear account of what consumes credits, distinguishing record writes, list changes, external requests and integration blocks at one credit each from AI blocks covering classification, summarisation, web research and custom agents. That tells a buyer where the model is invoked, which is genuinely useful and more than most publish.

What it does not tell them is anything about the model itself: no provider, family or version is named, no model card exists, and no evaluation or accuracy figure is published for the assistant, the research agent or the classification attributes.

That last gap has a specific consequence in a system of record, because an AI attribute writes an inferred value into a field that then drives filters, views and workflow branching, and nothing published states how often those inferences are right. Third party analysis had to assume an average credit cost per AI run in the absence of a published rate. Ask which models process workspace data, and what accuracy the classification and research attributes achieve.

A large and consistent review base, and no measured outcome published anywhere. Independent review aggregation records roughly two hundred and twenty ratings at four point four out of five with around eighty six percent of reviewers at small business scale, which is a reliable signal of who uses the product and how they find it. Reviewers consistently praise the flexible data model and the interface, and consistently note reporting limitations at larger scale.

What is absent is the vendor's side: no published study measures adoption, data quality, time saved or revenue effect across a stated customer population and period, and no case studies with quantified outcomes were located. Third party analysis raises a cost of ownership point the vendor does not address, arguing that a flexible platform still depends on manual record maintenance and that the labour cost of keeping it current can exceed the licence. Ask for measured adoption and data completeness across a stated customer population, and for the vendor's own position on maintenance effort.

The platform sends, and none of the obligations that come with sending are addressed. Email sequences are available at the upper paid tier, mail leaves through user mailboxes connected by synchronisation, and per seat sending volumes are capped by plan, with the free tier limited to a stated monthly figure. Those caps are a volume control rather than a compliance one.

Nothing published describes suppression list handling, unsubscribe mechanics, how an objection recorded against one record propagates across sequences, or what happens when a contact added by automatic enrichment is then sequenced. That last case is the specific exposure in this product: records can arrive by enrichment rather than by the customer entering them, and can then be sequenced, so a person may receive outreach from a company that never knowingly collected their details. Ask how unsubscribe and suppression state is maintained across sequences, and what governs sequencing a contact that arrived through automatic enrichment.

The core instruments are published rather than gated, which is the right shape for this category. A privacy policy and a data processing addendum are both public, the addendum states the terms governing processing on customers behalf and identifies the subprocessors used, and the policy distinguishes where the vendor acts as controller from where it acts as processor for individuals in the European Economic Area, Switzerland and the United Kingdom.

That distinction is stated rather than blurred, which matters for a platform holding a customer's entire relationship history. A public trust centre carries further compliance documentation. Held below the top band because several elements were not established on this pass: no transfer framework certification was located, no retention period or deletion timeline is published, and no data protection officer or European representative was found, where vendors at the top of this axis publish request metrics, registration numbers and named representatives. Ask for the retention schedule, the deletion timeline after termination, and the transfer mechanism for European workspace data.

The platform enriches records automatically and does not say where the data comes from. Automatic enrichment is described as a feature of every tier including the free one, adding social and company information to records without the customer entering it, and enrichment consumes credits, so it is a metered service drawing on some external supply.

No provider, cooperative or licensing arrangement is named anywhere, no lawful basis is stated for the personal data being appended, and no indemnification position appears. This matters more in a system of record than in a prospecting tool, because enriched values sit permanently in the customer's own database alongside data they collected themselves, and after export the two are indistinguishable.

A buyer inherits whatever provenance the underlying source carries without being told what it is. Ask which providers supply enrichment data, on what lawful basis for European records, and whether provenance is indemnified.

Integration runs through sanctioned routes and the contract addresses circumvention directly. Mail and calendar connect through account synchronisation with a stated number of connected accounts per user by tier, identity runs on established standards with single sign on, security assertion markup and directory provisioning, and the workflow engine reaches external systems through published integration blocks and requests rather than through improvised means.

The customer agreement explicitly prohibits circumventing security or licence controls and prohibits pooling or multiplexing connections to reduce the number of users accessing the service, which is the vendor protecting its own seat model and incidentally describing a clear boundary. Nothing resembling scraping or unsanctioned automation appears.

Held below the top band because no stated conformance position against any specific platform's terms was located, and because integration blocks and the number of connectable mail accounts are both gated by tier, so the integration surface a buyer receives depends on what they pay. Ask which integration capabilities exist at the tier quoted, and whether any named platform conformance commitments exist contractually.

A system of record with AI reading across it, and no stewardship position located covering that combination. The assistant answers questions grounded in the workspace, research agents fetch external material into records, and classification and summarisation attributes read stored data, which means the customer's complete relationship history, including correspondence synchronised from connected mailboxes, is available to model infrastructure.

Nothing published states whether workspace data trains or tunes any model, which providers receive it, whether prompts and generated output are retained and for how long, or whether a customer can exclude particular objects from AI processing. A public trust centre exists and carries compliance documentation whose contents were not reachable on this pass, so this records what could be established rather than asserting no position exists. Ask whether workspace or synchronised mail data trains or tunes any model, which providers process it, and whether objects can be excluded from AI features.

Two surfaces reach people outside the customer and neither carries a stated position. Email sequences send to prospects from a seller's own connected mailbox, and the assistant can draft that content, so a recipient may receive machine written copy under a human's name with nothing indicating it.

Separately and more distinctively, automatic enrichment appends personal information to records for people who never gave it to the customer, and research agents fetch further material about them from the web into the same records. Nothing published describes what notice, if any, reaches those individuals, or how someone would discover that a company they have never contacted holds an enriched and machine researched profile of them.

The vendor does publish a privacy policy and processing addendum through which a rights request could in principle be made, which is why this sits mid band rather than lower. Ask what notice reaches individuals added by enrichment, and whether AI drafted sequence content is marked for the sender.

A programmable platform rather than a closed application, with the interface treated as a first class surface. A published application programming interface is available with developer documentation, integration blocks reach external services from inside workflows, and outbound requests are supported as a native workflow operation, which means a customer can extend the system without leaving it.

Identity and provisioning run on established standards including single sign on, security assertion markup and automated directory provisioning with a named identity provider. Mail and calendar synchronisation is bidirectional and real time, and a help centre documents configuration in genuine operational detail.

Held below the top band on gating and verification: integration blocks, the number of connectable mail accounts and single sign on are each tier restricted, so the entry and free tiers reach materially less than the published capability list suggests, and no integration inventory or rate limit documentation was verified on this pass. Ask which integration capabilities are available at the tier quoted, and for the interface rate limits.

Cloud hosted with no residency election published. Independent analysis describes the platform running on globally distributed cloud infrastructure for availability and performance, and observes that this same architecture constrains residency options for organisations required to keep data in a particular jurisdiction.

No hosting region is named on the vendor surface, no European or United Kingdom residency option is described, no tenancy model is stated, and no recovery time or recovery point objective is published. The gap is notable for a vendor headquartered in London selling substantially into European markets, where a buyer would reasonably expect a residency answer, and it is compounded by what the product holds: a complete relationship history including synchronised mail bodies and attachments at the upper tiers. Ask which regions workspace data is stored in, whether European or United Kingdom residency is available, and what the recovery objectives are.

The delivery model is better than almost anything in this index and the contents could not be read on this pass. The vendor publishes a trust centre and states plainly in its own help material that it is public, that no account is required, and that the link can be handed straight to a security or procurement team. It is described as carrying certifications, penetration test reports and security questionnaires.

Publishing penetration test reports and completed questionnaires without a confidentiality agreement is materially more open than the norm, where those are precisely the documents held behind an approved request. A privacy policy, a processing addendum naming subprocessors, and a customer agreement are all separately published, and the product supports single sign on, security assertion markup and automated provisioning. Two things hold it below the top band.

The certification set could not be read on the routes taken this pass, so which certifications are held is unconfirmed here rather than absent. And the vendor's own careers material describes a security leadership role whose remit is to build and lead the security function and drive compliance with named frameworks, which reads as a function being formalised rather than long established. Two competing vendors publish flatly contradictory claims about this vendor's attestation status and neither was used. Ask which certifications are current, with audit periods and auditors.

Among the strongest published pricing in this index, and short of the top band on two specific things. Published on the vendor's own page: four tiers with per user rates shown in both billing cycles, a free tier for up to three seats that is a real product rather than a trial, a twenty percent annual discount stated explicitly, and a fourteen day trial of the upper tier for new workspaces.

The gates that decide the real bill are enumerated rather than implied, with object counts, record ceilings, connected mail accounts per user, email sharing depth, permission levels and feature availability all laid out per tier. Overage is priced rather than left to a conversation, with credit packs published at three sizes in both billing cycles. A startup programme is published with stated discounts across two years. What holds it below the top band is the link between usage and cost.

The vendor publishes what consumes credits and what a pack costs, but not how many credits a given AI operation consumes, so a buyer can price a pack without being able to forecast how many they will need, and third party analysis has resorted to assuming an average. The comparable A grade in this category pairs its published tiers with equally prominent commercial terms covering cancellation, refunds and downgrade credits; no equivalent was located here. Ask for the credit consumption rate per AI operation type, and for the cancellation and refund terms.

The mechanism to leave exists and the terms of leaving are not published. A documented interface is available across paid tiers and would let a determined customer extract their records programmatically, which is a genuine portability route rather than a promise, though it requires technical work rather than a button.

Beyond that the public surface is silent on everything this axis asks: no statement of export scope or format, nothing on whether synchronised mail bodies, attachments, enriched values, workflow configuration and reporting history leave with the customer, no retention period after termination, no deletion timeline or confirmation, and no renewal or cancellation notice terms.

The category editorial identifies this axis as doing more work here than anywhere in the index, because a relationship platform holds years of history a customer cannot reproduce. Two specifics deserve an answer: what licence attaches to enriched values after termination, since those were sourced from third parties, and whether the flexible data model exports in a form another system can read. Ask what exports and in what format, what happens to enriched values and mail content, and what the deletion timeline is.

Mail leaves from the customer's own mailbox, which is the architecturally safer arrangement, and nothing beyond that is described. Sequences send through accounts connected by synchronisation rather than through vendor operated infrastructure, so sending reputation stays on the customer's own domain and there is no pooled estate where one customer's behaviour degrades another's delivery. That is a real structural property.

Per seat sending volumes are capped by tier, with the free tier limited to a stated monthly figure, which incidentally limits blast radius. What is absent is every operational control: nothing published on bounce handling, complaint thresholds, list validation before send, warmup for a newly connected mailbox, or what happens when a sequence begins returning failures.

For a product where records can arrive by automatic enrichment and then be sequenced, validation before send is the control that matters most and it is not described. Ask what bounce and complaint handling applies, whether addresses are validated before sequencing, and what happens when a mailbox starts failing.

The served segment is unusually legible, and the vendor puts money behind its stated target. A genuine free tier for up to three seats means the smallest possible team is served by product rather than by trial, and the ladder runs continuously from there through two published paid tiers to a custom enterprise arrangement.

Independent review data confirms where the platform actually lands, with roughly eighty six percent of reviewers at small business scale, and the vendor's positioning as an alternative to the heavier incumbents matches that. The startup programme is the sharpest signal: published discounts of eighty percent in the first year and thirty percent in the second for qualifying venture backed companies is a quantified commitment to a named segment rather than a marketing claim, and it tells a buyer plainly who this is priced for.

Held below the top band because the upper end is thinner than the lower: no industry pages or localisation were located, reviewers report reporting limitations at larger scale, and independent analysis describes the enterprise capabilities as comparatively recent. Ask what the largest reference deployment looks like in seats and record volume.