Email Deliverability Signals: How to Diagnose Problems, Use AI, and Make Better Sending Decisions

Email teams have more deliverability data than ever: inbox placement, sender reputation, SMTP responses, complaints, authentication, engagement, campaign activity, volume and mailbox-provider data. Yet when something suddenly goes wrong, the hardest questions are still the simplest ones.

• What happened?
• Why did it happen?
• What should we do next?

That was the focus of From Deliverability Signals to Sending Decisions: How AI Is Changing Email Operations, a Mailora webinar presented in association with the Canadian Email Summit. The conversation featured Chaitanya Chinta (CC), Head of Product at Halon Engage and a Mailora advisory board member, with Adeola Sole hosting.

Rather than another Deliverability 101 conversation about SPF, DKIM and DMARC, the session focused on what happens after the basics are already in place: how to investigate a decline, identify the traffic actually responsible, understand changing mailbox-provider behaviour, and decide when AI can help. This article brings together the strongest ideas from the session, along with the audience questions we received.

Delivery vs. deliverability: an important distinction

One of the first distinctions CC drew was between email delivery and email deliverability. They are used interchangeably, and they measure different things.

CC's example: an ecommerce brand sends a flash-sale email and the message reaches the inbox seven hours later. Technically it was delivered. Commercially it failed.

Delivery is an event. Deliverability is the quality of the outcome. Chaitanya Chinta, Halon Engage

The four signal groups behind email deliverability

CC grouped the factors affecting deliverability into four broad areas. No single one of them decides inbox placement on its own: mailbox providers watch all four dimensions of sender behaviour at the same time.

More email does not automatically mean worse deliverability

Adeola shared a real ecommerce example. Sending frequency increased substantially. Gmail stayed relatively stable, while Outlook and Yahoo behaved differently and delivery latency increased. The easy conclusion is "we sent more email, so our deliverability got worse".

CC's reading was more precise. Raw volume was not necessarily the problem. The change relative to the sender's historical pattern was.

A mailbox provider grows accustomed to a set of things about you:

• a certain daily volume
• a particular sending frequency
• a consistent level of recipient exposure
• reasonably predictable traffic

Change several of those quickly and the provider becomes more cautious. That often shows up first as throttling or increased delivery latency rather than an outright rejection.

Volume is a number. Behaviour is a pattern. Mailbox providers react to changes in the pattern.

Google's own sender guidance points the same way: keep sending rates consistent, avoid bursts, increase volume gradually, and monitor delivery as you scale (Google Help). Scaling email volume is not inherently bad. Uncontrolled scaling is the risk.

A practical troubleshooting framework: where, when, what?

The most useful framework in the webinar was CC's three-question approach. Work through them in order before changing anything.

01Where is the problem?

Determine the blast radius. A global campaign average hides the answer, so check whether the decline is:

• everywhere, or isolated to one provider such as Microsoft or Gmail
• confined to one IP, one domain or one campaign
• confined to one audience segment
• concentrated among newly acquired subscribers
• concentrated among older or less-engaged users

Gmail might be healthy while Microsoft deteriorates. In that case you do not have a deliverability problem. You have a Microsoft-specific deliverability problem, and that changes what you investigate next.

02When did it start?

Establish the timeline. Did performance deteriorate:

• immediately when the campaign began
• halfway through the campaign
• gradually across several sends
• immediately after a specific operational change

Timing narrows the hypothesis. A campaign that starts normally and deteriorates halfway through raises different questions from one that performs poorly from the first message: you may be looking at recipient reactions, traffic changes or filtering behaviour that developed during the send.

03What changed?

Compare the period immediately before the problem against your normal baseline. Ask whether any of these changed:

• volume
• frequency
• the audience
• the acquisition source
• the content
• the sending domain
• the infrastructure

The goal is not to prove causation immediately. It is to reduce the number of plausible explanations.

That last step matters. Teams often jump straight from "something changed" to "stop sending". There should be a diagnosis in between.

Stop asking "is this metric good?" Start asking "did this metric change?"

Industry thresholds have value. Gmail advises bulk senders to keep user-reported spam rates below 0.1% and never to reach 0.3%. But a threshold does not describe your sending program.

Suppose your complaint rate normally sits at 0.02% and suddenly reads 0.09%. You are still under the 0.1% guardrail, and something has changed by more than four times. That is worth investigating.

A useful sender baseline covers eight things:

CC's recommendation was direct: build the baseline, then compare new campaigns and operational changes against it, rather than reducing deliverability to a red or green status.

Delivery latency can be an early warning signal

Most senders notice a deliverability problem late, once inbox placement has dropped, complaints have risen, revenue has fallen or messages are visibly landing in spam. Increased temporary failures and progressively slower acceptance appear earlier than any of those.

If messages that normally arrive quickly suddenly take much longer at one provider, that does not prove a reputation problem on its own.

It does tell you the receiving system is treating your traffic differently than it normally does.

That is worth investigating before the problem grows.

Inbox placement is evidence, not absolute truth

Seed testing and inbox-placement monitoring give teams visibility that ESP reporting alone does not. CC added an important qualification: seed placement is not a perfect representation of every real subscriber's mailbox, and should be treated as directional external evidence.

It is most useful combined with:

• provider-specific delivery data
• reputation trends
• SMTP responses
• complaints
• campaign history
• changes against your normal baseline

The stronger diagnosis comes from connecting signals, not from elevating a single metric to the status of truth.

Where AI can actually help deliverability teams

Uploading one CSV to a language model and asking "what happened?" is the shallow end. The useful application is bringing the relevant sources together so AI can work across all of them at once:

• historical delivery logs
• mailbox-provider responses
• campaign activity
• reputation trends
• inbox-placement results
• other external signals

With that in place, AI can answer the questions a human would otherwise chase across five dashboards:

• What changed?
• Where did it change?
• When did it change?
• Which explanations are plausible?
• What should we investigate next?

The aim is not another AI-generated summary. It is hypothesis generation and investigation assistance.

But should AI make the final decision?

Not blindly. This was one of the clearest answers of the session. AI will tell you "this is the likely cause" and "this is the likely remediation" — and likely is the operative word. CC cautioned explicitly against implementing recommendations without review, because AI may not understand the full context of your business, reputation, data or commercial situation.

The rule of thumb: the higher the blast radius, the uncertainty, or the cost of reversing the decision, the stronger the case for human review.

When should you call a deliverability expert?

Not every problem needs an outside consultant. The webinar offered a clear dividing line.

The point is not to make deliverability mysterious enough that every issue needs a specialist. It is to recognise when uncertainty plus impact makes guessing expensive.

Eight deliverability questions marketers asked us

The live audience brought several strong questions. Some were answered during the session, others followed up afterwards. Responses are edited for clarity, with the substance of the speaker's answers preserved.

01How do you separate correlation from the real cause?

Start by building the timeline, then narrow the scope:

• Which providers were affected?
• Which campaigns?
• Which audiences?
• Which infrastructure?
• What changed immediately beforehand?

Where you can, change one variable at a time and look for repeated evidence. A reputation decline that follows a large campaign does not prove the campaign caused it. Treat it as a hypothesis to test, not a conclusion.

The sequence is correlation → hypothesis → evidence → intervention. Skipping the middle two steps is how teams end up fixing the wrong thing.

02What should marketers know about CASL in Canada?

The response highlighted three core requirements:

1. Obtain appropriate consent.
2. Clearly identify the sender.
3. Provide a functioning unsubscribe mechanism.

That matches current CRTC guidance on Canada's anti-spam legislation, which states that commercial electronic messages generally require consent, identification information and an unsubscribe mechanism, and that both express and implied consent exist depending on the circumstances (CRTC).

This is a summary of the webinar discussion, not legal advice. Assess your own CASL obligations with appropriate legal and compliance guidance.

03Can marketers still segment under GDPR and French privacy rules?

One clarification from the response is worth stating plainly: CNIL is France's data-protection regulator, not the regulation itself.

From a practical marketing point of view, two questions settle most cases:

• Do we genuinely need this attribute for this segmentation?
• Is using it consistent with what the person was told when the data was collected?

CNIL's own guidance emphasises purpose limitation, that personal data should be collected for a specified, explicit and legitimate purpose and not later reused in a way incompatible with it, alongside data minimisation (CNIL). Specific legal interpretations should go through privacy counsel.

04Do mailbox providers treat AI-generated email differently?

The practical risk is not "will Gmail detect that a model wrote this sentence?" It is "what behaviour is AI helping us scale?"

AI can help a team produce more relevant emails, better experiments and faster personalisation. It can equally help produce more irrelevant content, excessive frequency, poorly reviewed messaging and unwanted email at greater scale. Gmail's public sender requirements cover authentication, DNS, TLS, message formatting, spam rates, unsubscribe mechanisms and responsible sending behaviour, with no separate rule for AI-authored copy (Google Help).

AI does not remove the old rules. It accelerates whatever sending behaviour you already have.

05Will deliverability eventually become AI against AI?

Mailbox providers have relied on automated reputation and filtering systems for years. What is changing is the sender side, where AI increasingly helps teams interpret what receiving systems are signalling. A simplified version of the loop:

Receiving AI detects → sender AI diagnoses → human or policy decides → sending system adapts → receiving AI reacts.

This should not become a race to beat mailbox-provider AI. The useful goal is to understand recipient and provider feedback faster, and adapt sending practices responsibly.

06What skills should marketers develop as AI automates operations?

This may be the most important career takeaway from the session. If AI can analyse data, summarise reports, identify anomalies, produce campaign variations and automate operational tasks, then knowing which buttons to press stops being a differentiator. The valuable skills move upward:

• problem framing
• customer understanding
• commercial judgment
• experiment design
• data interpretation
• risk assessment
• strategy

CC's point was that AI can take on more of the operational analysis while the strategic layer stays with people who understand the audience, the brand and the business. The future-ready marketer is not the one who refuses AI. It is the one who knows when its output is useful, when it is wrong, and what to do with it.

07How should marketers approach Microsoft's spam filtering?

CC recommended treating Outlook as a distinct part of the sending strategy, with more deliberate audience segmentation where necessary. A useful diagnostic sequence:

1. Is the problem actually isolated to Microsoft?
2. Which audience segments are affected?
3. Did volume, frequency or engagement change?
4. Introduce remediation gradually and monitor the response.

The follow-up response added an important separation: Outlook.com consumer traffic and Microsoft 365 corporate environments are not one identical filtering environment, and should not be treated as one.

Provider-specific problems deserve provider-specific analysis.

08Which decisions can we trust AI to make today?

AI is increasingly capable of finding anomalous traffic, unusual SMTP-response clusters, sudden provider-specific changes, historical deviations, patterns across large datasets, and the campaigns or segments likely to be affected. That is valuable precisely because those tasks span millions of records nobody can inspect by hand.

But deciding whether to pause a campaign, move an IP, isolate a customer, migrate infrastructure or sharply reduce sending requires context AI may not have.

Use AI to reduce uncertainty. Do not let it hide uncertainty.

When the cost of being wrong is high, human judgment belongs in the loop.

What this conversation tells us about modern deliverability

One theme runs through almost every question in the webinar. Email teams do not suffer from a lack of data. They suffer from a lack of connection between the data.

A marketer may already have ESP campaign reporting, Google Postmaster data, sender reputation, authentication reports, SMTP responses, inbox-placement testing, complaint data, engagement data and campaign history. The next competitive advantage is not another isolated metric. It is being able to answer five questions:

• What changed?
• Where did it change?
• Why is it likely happening?
• How confident are we?
• What is the smallest safe action we can take?

That is the shift from deliverability monitoring to deliverability intelligence, and it is where AI is genuinely useful: not as a dashboard summariser, but as an investigation layer connecting signals a person would otherwise piece together by hand.

The final takeaway: better deliverability starts with better decisions

The best deliverability program is not the one with the most dashboards. It is the one that detects meaningful changes early, understands their context, and acts without disrupting healthy traffic.

• Build your baseline.
• Watch the changes.
• Diagnose the blast radius.
• Separate correlation from causation.
• Use AI to investigate faster.
• Keep humans responsible for the decisions where context and consequences matter most.

That is how a team moves from "something went wrong" to "we understand what changed, and we know what to do next."

Watch the complete webinar

From Deliverability Signals to Sending Decisions: How AI Is Changing Email Operations, featuring Chaitanya Chinta, hosted by Adeola Sole, presented by Mailora in association with the Canadian Email Summit.

Watch the webinar → Test your email with Mailora →